Speed up norm stats by skipping video decoding

[noobyalan]

Summary

Normalization statistics only use states and actions, but compute_norm_stats.py currently reuses the training data pipeline and decodes every video observation. This adds unnecessary CPU, memory, and I/O overhead, especially for high-resolution, multi-camera LeRobot datasets.

This PR adds an optional --skip-video-decoding flag for LeRobot datasets.

When enabled, it:

• keeps the existing dataset and transform pipeline;
• avoids downloading video files;
• replaces decoded frames with 1x1 uint8 placeholders;
• continues computing statistics from the original states and actions.

The option is disabled by default, so existing training and normalization behavior is unchanged.

Usage

uv run scripts/compute_norm_stats.py \
  --config-name <config_name> \
  --skip-video-decoding

Benchmark

Tested on a 1080p, three-camera LeRobot dataset:

Dataset	Frames	Default	Skip video decoding	Speedup
Subset	10,000	748.4 s	223.2 s	3.35×
Full dataset	159,423	7,629.3 s	403.6 s	18.9×

Statistics equivalence

The generated normalization statistics were identical between the default and optimized paths.

Testing

• Added a unit test covering disabled video downloads and dummy frame shape/dtype.
• Targeted data loader tests pass.
• ruff check, ruff format --check, and pre-commit pass on the modified files.

Related to #688.

[noobyalan]

@kvablack Hi! A gentle ping on this PR, it skips redundant video decoding in compute_norm_stats to speed up preprocessing, with unit tests included. Thanks!