[python] Import LeRobot video datasets

[XiaoHongbo-Hope]

Purpose

Extend load_from_lerobot to import LeRobot Dataset v3 video features into packed Paimon video fields, on top of the aligned multi-video file groups from #9480.

Changes

• infer all video features from info.json and configure video-frame-field
• map each Episode frame through the official from_timestamp + timestamp semantics
• keep every Episode in one aligned file group
• when the normal file or any camera sidecar needs to roll, close all active writers together at the next Episode boundary
• support nested source-camera MP4 boundaries without decoding or transcoding video
• read local, Hugging Face, and FileIO URI sources
• require an unpartitioned, bucket-unaware target so one Episode is handled by one writer
• keep source FileIO credentials isolated from the target Catalog

Encoded MP4 payloads are copied into packed video files and referenced by frame ordinal; they are not repeated once per frame. Payload deduplication is file-local, so a source MP4 shared across an aligned output-file boundary may be copied once into each output group. Per-Episode remux can remove that amplification in a follow-up.

Tests

• 233 related Python tests passed, 5 skipped, with 76 parameterized subtests
• official LeRobot 0.4.4 local Dataset integration: 20 passed
• shared-MP4 Episode rolling and oversized Episodes
• aligned group rolling triggered by one camera sidecar
• local and FileIO multi-video import with two Episodes and nested camera boundaries
• partitioned and fixed-bucket targets are rejected
• real lerobot/droid_100 metadata smoke: 3 cameras, Episodes 0 and 1 mapped to frame ordinals 0-165 and 166-403
• flake8, py_compile, and git diff --check passed