[videoAPI] fast seek implementation

[bjuncek]

🚀 Feature

Seek that seeks to the closest keyframe before requested timestamp

Motivation

At the moment, seek function is restricted to a precise seek which under the hood finds the closest keyframe, and decodes frames until we reach our targeted frame. This is a convenient, however potentially slower operation. We want users to be able to decide if they want to seek to keyframes only (for the fastest possible version), we should allow it for them.

Pitch

Imagine we have a video with keyframes at 0s and 8.3s. Then the precise seek (current implementation) works as follows:

video.seek(2)
data = next(video)
print(data['pts'])
>> 2.002

video.seek(5)
data = next(video)
print(data['pts'])
>> 5.005

video.seek(9)
data = next(video)
print(data['pts'])
>> 9.009

And the imprecise seek (keyframes only):

video.seek(2, keyframes_only=True)
data = next(video)
print(data['pts'])
>> 0.0

video.seek(5, keyframes_only=True)
data = next(video)
print(data['pts'])
>> 0.0

video.seek(9, keyframes_only=True)
data = next(video)
print(data['pts'])
>> 8.341667

Additional context

This is a function provided in pyav already, so I think it would be useful to have it as an option.

Note:

This requires changes in the decoder C++ API which is IIUC used internally in FB; before accepting the PR, @fmassa should probably look into importing it and testing it internally. The default values and operations are kept the same, but nevertheless, there could be regressions not caught by our current tests.

cc @bjuncek

[bjuncek]

closed via #3179