feat: add LeRobot v3 metadata validation checks

[kck325]

Summary

Add LerobotV3MetadataChecker — 7 pre-ingestion validation checks for LeRobot v3 datasets:

1. tasks.parquet — must exist (flags tasks.jsonl-only datasets)
2. Episodes parquet columns — requires data/chunk_index, data/file_index, tasks
3. Feature shapes — no empty shapes ([] should be [1])
4. Path templates — data_path and video_path must use standard placeholders ({episode_chunk}, {episode_index}, {video_key})
5. Video file existence — all files referenced by episodes must exist
6. Timestamp consistency — all relative or all absolute across data parquet files
7. Episode contiguity — rows for each episode_index must be contiguous

Motivation

These checks catch data issues at sharing time rather than at conversion time, preventing runtime fixups in downstream consumers. Partners can validate their datasets before uploading.

Test plan

• 17 new tests covering all 7 checks (pass and fail cases)
• Updated existing test fixtures to include v3 metadata
• All 51 tests pass with zero regressions

🤖 Generated with Claude Code