Remove stale external-overview skips on dask/fsspec/GPU golden-corpus backends

[brendancol]

Closes #3024

What

The golden-corpus dask, fsspec, GPU, and dask+GPU parity modules listed overview_external_ovr_uint16 in _OVERVIEW_READER_GAPS, which switched off the overview-level oracle comparison and ran only the base-IFD check. The reader has external .ovr sidecar discovery (_sidecar.py) shared across backends, so these gates were stale.

• Removed the overview_external_ovr_uint16 entry from _OVERVIEW_READER_GAPS in all four modules. The overview-level comparison now runs and passes.
• For fsspec, also pushed the sibling .tif.ovr into the memory:// filesystem so the reader's fsspec sidecar discovery (_probe_fsspec -> load_sidecar) can resolve the external overview levels. Without it the memory filesystem held only the base .tif and overview_level>=1 failed as out of range. That was a test-harness gap, not a reader limitation.

Backends

Test-only change. All four backends (dask+numpy, fsspec, GPU/cupy, dask+GPU) run and pass the overview comparison in the dev environment used here, which has a working CUDA device. Eager numpy already had its gap cleared.

Test plan

• test_dask_numpy_parity[overview_external_ovr_uint16] passes
• test_fsspec_parity[overview_external_ovr_uint16] passes
• test_gpu_parity[overview_external_ovr_uint16] passes
• test_dask_gpu_parity[overview_external_ovr_uint16] passes
• Full golden-corpus suite for the four modules: 127 passed, 9 skipped
• test_taxonomy_ids_are_in_manifest passes in each module

[brendancol]

PR Review: Remove stale external-overview skips on dask/fsspec/GPU golden-corpus backends

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• test_fsspec.py: the sidecar is served inside test_fsspec_parity via the memory_fs_clean fixture handle, while the base .tif goes through the module-level _serve_via_memory helper. Both write to the same process-global memory filesystem so it works, but a reader scanning the test sees two write paths for related files. Folding the sidecar write into _serve_via_memory (or a small sibling helper) keeps the memory-FS setup in one place. Cosmetic; the current form is correct.

What looks good

• The four skip removals are verified rather than assumed. Each backend's overview-level comparison was run and passes, and the taxonomy guard tests still pass because the entries were removed cleanly, not left dangling.
• The fsspec change is the substantive one and is right. That skip was masking a test-harness gap (only the base .tif was in the memory filesystem), not a reader limitation. Serving the sibling .tif.ovr lets the real fsspec sidecar discovery path (_probe_fsspec -> load_sidecar) resolve the external overview levels, so the test exercises production code instead of routing around it.
• The sidecar path construction (path.name + .ovr) matches find_sidecar's source + ".ovr" convention, so the served URL lines up with what the reader probes.
• Comments left in place of the removed entries explain why the gap is gone, matching the eager numpy module's existing note.

Checklist

• Algorithm matches reference/paper (n/a, test-only)
• All implemented backends produce consistent results (all four verified passing)
• NaN handling is correct (n/a)
• Edge cases covered by tests (this PR adds the previously-skipped overview coverage)
• Dask chunk boundaries handled correctly (n/a, no production change)
• No premature materialization or unnecessary copies (n/a)
• Benchmark exists or is not needed (not needed, test-only)
• README feature matrix updated (n/a, no new function)
• Docstrings present and accurate (skip-list comments updated)

[brendancol]

Follow-up review (after edcf7bb)

The single nit from the first pass is resolved. The fsspec sidecar write now lives in _serve_via_memory alongside the base .tif write, so all memory-filesystem setup is in one helper and the test body no longer carries a parallel write path. Full test_fsspec.py module passes (32 passed, 2 skipped).

No new findings. Nothing outstanding.