bench: run the full compress suite under --gpu-decompress

[joseph-isaacs]

Rationale for this change

--gpu-decompress in compress-bench was gated behind an incremental allow-list holding a
single entry (TPC-H l_comment canonical), so the GPU compression benchmark only ever measured
one dataset. Widening it to the whole suite surfaced six independent gaps in the GPU decode path,
each fixed here. The benchmark now runs all 17 datasets end to end on the GPU.

What changes are included in this PR?

Benchmark scope

• Remove the gpu_decompress_benchmarks allow-list. --gpu-decompress now runs every dataset in
  the suite: taxi, the Public BI sets, both TPC-H l_comment variants, the StructListOfInts
  wide-table variants, plus airquality and rplace.
• Install the DuckDB CLI in the GPU benchmark workflow. The Public BI CSV-to-parquet conversion
  shells out to it, and the CPU benchmark runner already did this.

GPU decode gaps found by the widened suite

• DateTimeParts dispatched through match_each_signed_integer_ptype!, panicking with
  "Unsupported ptype u16" on taxi. Compression picks the narrowest ptype per component and the
  CPU decoder already accepts any integer ptype, so match that and generate the kernel for every
  signed and unsigned width.
• RunEnd bailed when values carried Validity::Array. The error claimed a CPU fallback, but
  execute_cuda can only fall back while all buffers are host-resident, so a device-resident scan
  failed outright. New runend_bool kernel expands the per-run validity bitmap through the run
  mapping, one output byte per thread.
• No CUDA kernel for vortex.masked. New MaskedExecutor: a MaskedArray is a child carrying no
  nulls plus the bitmap that supplies them, so decode both on the GPU and attach the mask.
• No CUDA kernel for vortex.list. New ListExecutor and list_views kernel deriving the
  ListView offset/size pair from List's len + 1 Arrow offsets.
• only_cuda_compatible() excluded the integer and float sparse schemes but missed
  string::NullDominatedSparseScheme, so the CUDA-compatible compressor still emitted
  vortex.sparse (which has no GPU decoder) for string columns.
• into_host copied data buffers off the device but left Validity::Array bitmaps on it
  ("unwrap_host called for Device allocation"), and had no Canonical::List arm at all.

Tests

• New coverage for unsigned DateTimeParts components, RunEnd validity expansion across byte
  boundaries, masked validity, and list decode (contiguous, empty, and nullable).
• The nullable RunEnd tests now call the executor directly. Going through execute_cuda silently
  falls back to CPU for a host-resident array, so they passed without ever running the GPU kernel.

What APIs are changed? Are there any user-facing changes?

No public library API changes. compress-bench --gpu-decompress no longer filters the dataset
list, and three encodings (vortex.masked, vortex.list, and RunEnd with per-element validity)
now decode on the GPU instead of erroring.

Checks

• cargo clippy -p vortex-cuda -p vortex-btrblocks -p compress-bench --all-targets, cargo +nightly fmt --all, cargo test -p vortex-btrblocks --lib builder, and
  yamllint --strict all pass locally.
• CI CUDA tests (plus memcheck, initcheck, synccheck) pass on real hardware.
• The GPU compression benchmark completes all 17 datasets.

One thing worth a reviewer's judgement: the GPU path now downloads airquality and rplace from
pcodec's public S3 bucket on every run. The CPU-side exclusion of those two exists to avoid
creating egress charges for pcodec, so this may want caching rather than a standing CI cost.

[codspeed-hq]

Merging this PR will degrade performance by 1.6%

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

⚡ 1 improved benchmark
❌ 1 regressed benchmark
✅ 2035 untouched benchmarks
⏩ 46 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	WallTime	words_gather_scalar[65536]	8.3 µs	9.4 µs	-11.98%
⚡	WallTime	words_gather_dispatch[1024]	33 ns	30 ns	+10%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing claude/gpu-compress-benchmarks-x788xc (8d6595e) with develop (1a493b6)}

Footnotes

1. 46 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[github-actions]

GPU Compression

┌───────────────────────────────────────────────────────────┬────────────────────────┐
│ Benchmark                                                 │ vortex-file-compressed │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/taxi                                      │ 109761.00 μs (1.00)    │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/Arade                                     │ 548279.00 μs (1.00)    │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/Bimbo                                     │ 1428127.00 μs (1.00)   │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/CMSprovider                               │ 6595018.00 μs (1.00)   │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/Euro2016                                  │ 928742.00 μs (1.00)    │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/Food                                      │ 354424.00 μs (1.00)    │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/HashTags                                  │ 2120291.00 μs (1.00)   │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/TPC-H l_comment chunked                   │ 1963364.00 μs (1.00)   │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/TPC-H l_comment canonical                 │ 1956508.00 μs (1.00)   │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/rplace                                    │ 2597239.00 μs (1.00)   │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/airquality                                │ 21991.00 μs (1.00)     │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/wide table cols=100 chunks=1 rows=1000    │ 7089.00 μs (1.00)      │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/wide table cols=1000 chunks=1 rows=1000   │ 59950.00 μs (1.00)     │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/wide table cols=10000 chunks=1 rows=1000  │ 645809.00 μs (1.00)    │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/wide table cols=100 chunks=50 rows=1000   │ 7019.00 μs (1.00)      │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/wide table cols=1000 chunks=50 rows=1000  │ 61531.00 μs (1.00)     │
├───────────────────────────────────────────────────────────┼────────────────────────┤
│ decompress time/wide table cols=10000 chunks=50 rows=1000 │ 639750.00 μs (1.00)    │
└───────────────────────────────────────────────────────────┴────────────────────────┘