[python] Support video frames in contiguous window datasets

[YannByron]

Summary

ContiguousWindowDataset currently rejects video-frame-field columns, so callers cannot train directly from video-backed windows. This change keeps video descriptors through row selection and decodes the selected frames only after overlapping window reads have been coalesced.

Changes

• Decode video-frame-field columns in ContiguousWindowDataset with the existing PyAV and TorchCodec decoder path.
• Group rows by physical video and decode each requested frame once per batched read while preserving window order.
• Keep ordinary BLOB body fetching unchanged and route indexed video range I/O through the existing blob_parallelism setting.
• Add video_backend selection to ScanQuery.to_contiguous_window_dataset.
• Document direct N x T x C x H x W DataLoader training usage and backend dependencies.

Testing

• pytest -q pypaimon/tests/contiguous_window_dataset_test.py pypaimon/tests/multimodal_video_test.py (54 passed)
• pytest -q pypaimon/tests/multimodal_lerobot_test.py -k 'video_decoder or pyav or indexed' (7 passed)
• Real MP4 encode/decode through a two-worker PyTorch DataLoader
• flake8 --config=./dev/cfg.ini on changed Python files
• Python license header check (960 files)

[JingsongLi]

Found one reproducible memory issue in the video decoding path; details inline.

[JingsongLi]

[P2] Give each decoded frame independent storage

get_frames_at(...).data is a batched Tensor, so the collator assigns views of its storage to individual rows. _sample() then calls copy.deepcopy() separately for each requested frame (lines 303–305), which copies the entire backing batch storage for every view. With no column_transforms, a real PyAV read of 32 three-frame windows (34 decoded rows) returned 96 independent storages of 4,896 bytes each, although each 8×6 RGB frame contains only 144 bytes: 470,016 retained bytes instead of 13,824 (34×). torch.stack releases those oversized copies after each transform, but still incurs the full batch copies. This substantially increases memory and copying costs for video-training batches. Please clone each frame before passing it to the collator, or use Tensor.clone() instead of deepcopy in _sample(), and add a regression check for the returned frames' storage size. Per-frame cloning reduced this reproduction to 1× storage with identical frame values.