[python] Simplify LeRobot metadata tables and training tags - #9676

Purpose

Decouple LeRobot ingestion from dataset version publication. Store dataset info and optional statistics in companion tables with key STRING, value STRING, using JSON-encoded values to preserve nested metadata without a fixed schema.

load_from_lerobot returns None on success and accepts an optional tag_name. create_lerobot_tag pins the current snapshots of all recorded components under a shared name for training, checks conflicts before creating tags, and supports retries while snapshots remain unchanged.

Cross-table tagging is not atomic: callers must pause group writes until the call succeeds and read every component through the same tag. Documentation and regression tests cover this contract.

Tests

• LeRobot import suite: 54 tests and 46 subtests passed, including JSON round-trips, optional metadata, stable tagged reads after appends, conflicts, missing components, and partial tag failure/retry.
• Multimodal table and filesystem catalog tag suites: 103 passed, 2 skipped because Ray was unavailable.
• Flake8 on changed Python files and git diff --check passed.

LeRobot tests ran with Python 3.12, LeRobot 0.4.4, PyArrow 19.0.1, and NumPy 2.3.5. NumPy 2.5 fails in the existing SDK fixture's array-to-scalar conversion; project dependencies are unchanged.