[python] Add ACT storage benchmark and contiguous windows

[YannByron]

Summary

Add a lazy, snapshot-pinned contiguous-window Dataset and a reproducible ACT
storage benchmark for RoboMIND HDF5 and Paimon data.

HDF5 and Paimon runs are independent commands. They consume one resolved
experiment containing the same configuration, normalization, seed, episode
selection, snapshot, and logical window sequence. Comparison verifies that
contract, runtime environment, tensors, and loss traces before reporting
performance ratios.

Changes

• Add ContiguousWindowDataset with deterministic group/order indexing,
  row-ID reads, snapshot pinning, tail policies, anchor-only projections,
  transforms/adapters, and coalesced plural __getitems__ access.
• Resolve scalar BLOB and MAP<..., BLOB> values lazily, isolate mutable
  values between overlapping windows, and normalize conflicting scan options
  before snapshot reads.
• Add prepare, run, and compare ACT commands with separate HDF5 and
  Paimon adapters and one shared model/training harness.
• Package default experiment parameters and write the complete resolved
  experiment into every result artifact.
• Record dependency/build versions, CPU identity/count, Torch thread settings,
  source identity, tensor fingerprints, losses, timing, and Python allocation
  metrics. Batch-fetch timing excludes tensor validation.
• Document the window API, benchmark workflow, result comparison, measurement
  boundaries, and Python 3.10 requirement.

Testing

• Rebased onto origin/master at 8d67f8c6f.
• python -m pytest pypaimon/tests/contiguous_window_dataset_test.py pypaimon/tests/act_benchmark_test.py pypaimon/tests/act_runner_test.py -q (42 passed).
• python -m pytest pypaimon/acceptance -q (3 passed).
• Reviewer regressions cover map-only and mixed scalar/map BLOBs with
  null/empty values, mutable overlapping/duplicate windows, latest-full scan
  pinning, fetch-only timing, runtime identity, and the Python version floor.
• Ruff and Flake8 on the changed Python files.
• Python license check (738 files have valid license headers).
• python -m compileall -q pypaimon/benchmark/act pypaimon/multimodal.
• git diff --check.
• Built and inspected the wheel for ACT modules, the default experiment,
  the window Dataset, and guarded optional dependencies.

A raw local python -m pytest -q run completed with 4691 passed, 137 skipped. Its remaining failures were environment-specific: one missing
DataFusion dependency passed after installing the CI version, the macOS
torchrun test timed out, and Java interoperability fixtures were not built.

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 not validated by this local CPU benchmark.
• This adds Python APIs and benchmark tooling; it does not change an existing
  public API.
• This builds on the RoboMIND AgileX HDF5 pipeline merged in [python] Add RoboMIND AgileX HDF5 pipeline #9445.