Feat/norm stats compatibility report

[TheRealBaoBao]

Summary

Adds a read-only CLI that compares post-transform training samples with the normalization statistics selected by an OpenPI training config.

The report identifies:

• Data/stat dimension mismatches
• NaN and Inf values
• Near-constant dimensions
• The fraction of values outside configured q01–q99 bounds
• The dimension with the largest mismatch

Example:

uv run scripts/check_norm_stats.py \
  --config-name pi05_droid_finetune \
  --max-frames 5000

The command reuses the existing config and normalization-stat data loaders. It does not modify the dataset, statistics, model, configuration, or training path.

Motivation

When adapting π0 or π0.5 to a custom robot dataset, reused checkpoint statistics can be structurally valid but poorly matched to the transformed state/action distribution. This report makes that mismatch visible before a long fine-tuning run.

This is related to the normalization questions and failures discussed in #773, #814, and #832.

Tests

• uv run pytest -q src/openpi/training/norm_stats_report_test.py — 7 passed
• uv run ruff check scripts/check_norm_stats.py src/openpi/training/norm_stats_report.py src/openpi/training/norm_stats_report_test.py
• uv run pre-commit run --files scripts/check_norm_stats.py src/openpi/training/norm_stats_report.py src/openpi/training/norm_stats_report_test.py
• uv run scripts/check_norm_stats.py --help

Scope

This intentionally excludes dataset mutation, automatic normalization-stat selection, general dataset-quality checks, and changes to normalization or training behavior.

I would appreciate feedback on whether maintainers prefer this as a separate CLI or as an optional mode of compute_norm_stats.py.