[python] Read LeRobot video datasets for training

[XiaoHongbo-Hope]

Purpose

Read video-backed LeRobot v3 table groups through PaimonLeRobotDataset.

Changes

• decode .video frame descriptors with TorchCodec or PyAV
• reuse bounded decoder caches per DataLoader worker
• build a lightweight packet timestamp/keyframe index for cold PyAV random reads, then seek directly to the target GOP
• support existing Episode, delta-window, padding, transform, and return_uint8 behavior

TorchCodec is preferred; PyAV is the fallback.

Tests

83 LeRobot/video tests passed, including real MP4, partial delta windows, B-frame random rereads, cold random-read decode bounds, optional-dependency gates, and spawn workers.

[JingsongLi]

+1