Video support

[lhoestq]

(wip and experimental)

adding the Video type based on VideoReader from decord

>>>from datasets import load_dataset
>>> ds = load_dataset("path/to/videos", split="train").with_format("torch")
>>> print(ds[0]["video"])
<decord.video_reader.VideoReader object at 0x337a47910>
>>> print(ds[0]["video"][0])
tensor([[[73, 73, 73],
         [73, 73, 73],
         [73, 73, 73],
         ...,
         [23, 23, 23],
         [23, 23, 23],
         [23, 23, 23]]], dtype=torch.uint8)

the storage is the same as for audio and images: {"path": pa.string(), "bytes": pa.binary()} and I did a small to keep the hf:// URL in the "path" field if possible, this way the viewer can link to fiels on the hub if possible

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.