[python] Add paired HDF5/Paimon ACT benchmark

[YannByron]

Summary

Add a paired ACT benchmark that compares the original RoboMIND AgileX HDF5 path with the Paimon path under one deterministic training contract. Both backends share the same configuration, train-only normalization, random seed, window sequence, ACT policy, optimizer, and correctness gates.

Changes

• Add a snapshot-pinned, lazy ContiguousWindowDataset for map-style PyTorch access to complete row windows.
• Coalesce overlapping rows during batched access and support anchor-only columns so observation BLOBs are loaded once per action window.
• Add HDF5 and Paimon ACT adapters plus a shared deterministic CPU harness with tensor and loss parity checks.
• Alternate backend execution for at least three rounds and report first-batch latency, DataLoader throughput, fixed-step time, and separately measured Python allocations.
• Reuse the RoboMIND pipeline's versioned canonical-action statistics and reject source, snapshot, normalization, or window-sequence mismatches.
• Add act and hdf5 optional dependencies, CLI guidance, and PyTorch/window API documentation.
• Fall back to UNKNOWN source provenance when the installed package is not inside a Git checkout.

Testing

• PR3 plus packed-video conflict regression (paired_act_benchmark, contiguous_window_dataset, RoboMIND, Ray HDF5, multimodal video/table, and Torch tests): 144 passed, 1 skipped
• Non-E2E Python suite before the latest master rebase (4523 passed, 134 skipped; one unrelated Ray shuffle ordering assertion failed once and passed on isolated rerun)
• Isolated Ray shuffle rerun (1 passed)
• Ruff on all changed Python files
• Python license checks (707 files valid)
• compileall and git diff --check
• Wheel build and content check for the benchmark and window-dataset modules
• Full ACT benchmark on a downloaded RoboMIND dataset (requires dataset and LeRobot model assets)

Notes

This builds directly on the RoboMIND AgileX HDF5 pipeline merged in #9445. Ingestion and canonical-action backfill are intentionally outside the measured benchmark scope.