Fix timestamps

[moi15moi]

This PR is a draft because i'm waiting for FFMS/ffms2#472 to be merged.
As we already talk (here and in gjm discord), mpv shift the video timestamps by AVFormatContext::start_time.
Fix #172

[moi15moi]

ffms2 case, I think time_of_frame_0 is always equals to AVStream::start_time (this might not be the case!!!)

Actually, yes and no.
If you use the sample in this issue with ffmpeg 8.0.1 or less, the first frame PTS is NOT the same as AVStream::start_time, but this was indeed a bug, so i'm not sure if that's relevant.

[arch1t3cht]

I'm fairly sure that mpv does not read AVStream::start_time anywhere. Judging by its code it just offsets all timestamps by -AVFormatContext::start_time.

[arch1t3cht]

Moreover, AVFormatContext::start_time seems to just be computed as (more or less) the minimum of the start_times of all of the file's tracks, see the update_stream_timings function in ffmpeg. So rebasing by start_time would, for example, not help with a 83ms delayed file that only contains a video track.

There's probably no good way to fix this, and the simplest solution is to just a) rebase by start_time, which at least matches mpv, and b) simply alert and ask the user whenever the first timestamp is not 0. I'm mostly ffmpeg's behavior here for future reference so I don't forget this later.