feat(action): publish action stats utility

[zichen0116]

Fixes #168.

Summary

This PR adds a public utility for generating action normalization statistics from local LeRobot-compatible datasets.

Cosmos action-policy recipes consume precomputed statistics files, but the generation script is not included in the public repository. This makes it difficult to recompute statistics for a new robot embodiment or after changing dataset preprocessing.

Changes

• Add tools/compute_action_stats.py.
• Support one or more local LeRobot dataset roots.
• Compute min, max, mean, std, q01, and q99.
• Use deterministic reservoir sampling so quantiles do not require loading the full dataset into memory.
• Validate action dimensions and reject empty, NaN, or infinite values.
• Emit the flat JSON schema consumed by the existing action normalizer.
• Add CPU-only tests for statistics, deterministic sampling, and invalid input handling.

Usage

python tools/compute_action_stats.py \
  --dataset-root /path/to/lerobot-dataset \
  --action-key action \
  --output stats.json \
  --reservoir-size 50000 \
  --seed 42

For multiple compatible datasets, pass multiple paths after --dataset-root.

Non-goals

• No changes to model architecture or default normalization behavior.
• No GPU requirement.
• No changes to existing bundled statistics files.

Testing

• pytest -o addopts= --confcutdir=tools tools/compute_action_stats_test.py — 6 passed
• ruff check --isolated tools/compute_action_stats.py tools/compute_action_stats_test.py
• ruff format --check --isolated tools/compute_action_stats.py tools/compute_action_stats_test.py
• python -m compileall -q tools/compute_action_stats.py tools/compute_action_stats_test.py