Add video support

[lhoestq]

videos in /first-rows and /rows will be like

{"src": "https://huggingface.co/datasets/username/dataset_name/resolve/commit/hash/path/to/video.mp4"}

since videos are stored on the hub as files in 99% cases :)

(otherwise the back-ends store videos in the assets on S3 as usual (e.g. if they come from parquet or webdataset data or zipped data) but this is NOT recommended)

In particular, the pinned datasets implements the Video type with decord and does NOT embed video bytes in Parquet files (HF URL only) when the file is on HF when preparing Parquet files for /rows.

Note that I also disable video decoding in get_rows() to not have to download the video data in /first-rows (HF URL only as well)

cc @AndreaFrancis for viz

[severo]

I remember you mentioned at one point that some video datasets have one video, and each row refers to a range [start, end], which would be accessed as "media fragments" in HTML: https://developer.mozilla.org/en-US/docs/Web/Media/Audio_and_video_delivery#specifying_playback_range. Is it still a thing, or do we just have one row = one video?

[lhoestq]

one row = one video (we'll see later if we need to support other cases)