feat: add P0 v3 validators for LeRobot dataset pre-ingestion checks

[kck325]

Summary

• Add 6 P0 validators as a new lerobot_validator/v3_checks.py module that catch the most common data quality issues before partner upload:
  • V1 validate_tasks_format -- error if neither tasks file exists, warn if only jsonl (old format)
  • V2 validate_codebase_version -- reject datasets without codebase_version starting with v3.
  • V5 validate_feature_shapes -- reject shape: [] (0-D), require 3-element shape for image/video features
  • V7 validate_timestamps -- reject absolute Unix epoch timestamps in data parquets; warn on non-monotonic or large starting offsets
  • V11 validate_custom_metadata_csv -- require episode_index/episode_id columns, reject null/duplicate episode_id values
  • V12 validate_start_timestamp -- require start_timestamp values are plausible Unix epoch floats (year 2000-2100 range)
• Wire validate_v3_dataset() into the LerobotDatasetValidator orchestrator so issues surface automatically
• Add get_warnings() method to the orchestrator and display warnings in print_results()
• Export Issue and validate_v3_dataset from __init__.py for direct usage
• Add 40 new tests in tests/test_v3_checks.py covering all validators and edge cases
• Update existing test fixtures in test_integration.py and test_is_eval_data_consistency.py to include codebase_version and tasks.parquet required by the new checks

Test plan

• All 40 new tests in test_v3_checks.py pass
• All 74 tests (including existing ones) pass with pytest tests/ -v
• Run validator against a real partner dataset to verify end-to-end behavior

🤖 Generated with Claude Code

[kck325]

Closing duplicate -- changes pushed to PR #8 on branch chandra/lerobot-v3-metadata-checker instead.