[python] Add paired ACT storage benchmark

[YannByron]

Summary

Add a reproducible paired ACT benchmark for equivalent RoboMIND data stored in HDF5 and Paimon.

This PR now contains only the benchmark. It builds on the snapshot-pinned ContiguousWindowDataset merged in #9580.

Changes

• Add prepare, run, and compare commands for independent HDF5 and Paimon runs using one resolved experiment.
• Verify source identity, logical window order, tensors, and loss traces before comparing performance metrics.
• Use frame_index consistently as the ACT sample position across both storage backends.
• Package the default experiment and optional ACT benchmark dependencies.
• Document the workflow, reported metrics, and measurement boundaries.

Testing

• Rebased onto the latest master (7d7a67281).
• python -m pytest pypaimon/tests/contiguous_window_dataset_test.py pypaimon/tests/act_benchmark_test.py pypaimon/tests/act_runner_test.py -q (51 passed).
• python -m pytest pypaimon/acceptance -q (3 passed).
• python -m flake8 --config=dev/cfg.ini pypaimon/benchmark/act pypaimon/tests/act_benchmark_test.py pypaimon/tests/act_runner_test.py setup.py.
• python dev/check_license_header.py (739 files have valid license headers).
• python -m compileall -q pypaimon/benchmark/act.
• git diff --check.

Notes

• Ingestion and canonical-action backfill are outside timed benchmark runs.
• The benchmark does not control the OS page cache.
• GPU, multi-worker loading, distributed training, recovery, and policy quality are outside this local CPU benchmark.

[JingsongLi]

Thanks for putting this together. This PR currently combines two independently valuable changes with different review and risk profiles: the generic ContiguousWindowDataset public API, and the RoboMIND ACT benchmark harness and workflow.

Given the 4.4k-line scope, could we split this into two PRs?

1. Contiguous-window Dataset: the public API, row-ID and snapshot semantics, BLOB handling, batching behavior, tests, and API documentation.
2. ACT benchmark: the prepare/run/compare workflow, backend adapters, packaging, benchmark documentation, and ideally a real paired Vortex result artifact with variance and physical-layout identity.

This would make the public API easier to review on its own, while allowing the benchmark methodology and reproducibility contract to be evaluated independently. It would also keep follow-up fixes in either area from blocking the other.

[JingsongLi]

Three actionable issues found in the updated benchmark-only scope.

[JingsongLi]

[P1] Compute both action statistics in the same numeric domain

The HDF5 moments are computed from the raw float64 actions, while the Paimon statistics were computed after the canonical actions were cast to float32. With this 1e-10 tolerance, ordinary decimal values can reject an otherwise equivalent dataset (for example, the mean of 0.1, 0.2, and 0.3 differs by about 5.5e-9 after a float32 round trip). The current integer-based test data masks this. Please cast the HDF5 actions to float32 before accumulating the moments, or compare the float32 normalization actually consumed, and add a decimal-valued regression case.

[JingsongLi]

[P1] Install Vortex for the documented workflow

The standard RoboMIND ingestion configuration uses vector.file.format=vortex, but this command installs only the act and hdf5 extras. Vortex is a separate extra, and its pinned package is available only on Python 3.11+, so a clean environment following these instructions cannot read the default warehouse; the stated Python 3.10 baseline cannot run that path at all. Please require Python 3.11+ and include vortex here (or make it part of the act extra), and cover the default Vortex format with a smoke test.

[JingsongLi]

[P2] Move tensor fingerprinting out of the pre-timing path

_tensor_fingerprint constructs the dataset and reads and decodes every distinct planned sample before any round starts. The subsequent first-batch and throughput measurements access those same windows, so the measured working set is deterministically warmed outside warmup_batches, and first_batch_s cannot represent a cold first access. The page-cache disclaimer does not cover this benchmark-induced pre-read. Please compute the fingerprint after all timed rounds or in an isolated validation process.