Initial support for video

[pzelasko]

This PR implements some features related to video dataloading support.

TL;DR

recording = Recording.from_file("example.mp4")
video, audio = recording.load_video()

cut = recording.to_cut().truncate(duration=1.0)
video, audio = cut.load_video()

More or less detailed list of changes:

• Recording
  • Recording.from_file() supports video files (so does lhotse.info())
  • Recording.load_video() method that loads video + audio (and keeps them in sync duration-wise)
  • the video shape is (num_frames, color, height, width), dtype is uint8, and format is RGB - if another makes more sense we can change it later
  • support only a single video stream for now (no multi-video files)
  • metadata about videos is in recording.video
  • check if recording.has_video
  • supports dynamic video resolution scaling with recording = recording.with_video_resolution()
  • can't load video when using perturb speed/tempo
• Cut
  • Supports all cut types (Mono, Multi, Mixed)
  • Mixed video cuts only support padding + appending with other video cuts, not mixing (but it's OK to mix audio-only cut into video +/- edge cases)
  • cut.has_video and cut.video (metadata)
• Dataloading
  • collate_video method that pads with black frames and packs video examples into 5d tensor
  • UnsupervisedAudioVideoDataset as a basic example for working with videos, returns video + audio mini-batches
• First supported audio-video corpus
  • Grid audio-visual speech corpus

For audio-only workflows, the code is otherwise practically unaffected. Dataloading videos with torch DataLoader generally works without issues. At this point though it's likely the current code won't work with every type of video format, and might not support some other standard Lhotse operations for video data yet. And it'll only work with recent PyTorch versions.