This removes all of the code that was previously used to get them from
/timedtext, and instead, always uses whatever is extracted from the
video page.
This does unfortunately now require a whole video fetch just for the
captions. But assuming captions are only requested by a frontend, this
won't be a problem due to the memory cache. The captions link will be
in memory because the just-requested video is in memory too.
Got some inspiration from https://github.com/nlitsme/youtube_tool (for the x-youtube-client-X
headers).
This is not a complete reimplementation of Invidious API as continuation is not implemented
(to retrieve more than the first 20 comments and comments replies), likes and replies count
are also missing.
The ads sections had a carouselAdRenderer property, now they have a
promotedSparklesTextSearchRenderer property instead. As this may
change again in the future, we should just get all items as we
discriminate/filter them as videos afterwards with the videoRenderer
property.
If error alerts exist, they will be logged. But it is reasonable to
assume that not all errors will be fatal, so we don't necessarily quit
parsing if we find one.
This also normalises the text error of the /latest response for a
missing channel, without changing its identifier.
Fix#27 use maintained yt-dlp lib instead of youtube-dlc
Because of the following changes in YT, we have to switch to a
maintained library https://github.com/ytdl-org/youtube-dl/issues/28604
While yt-dlp is not fixed today, youtube-dl is fixed in master and as
yt-dlp is quick to merge upstream changes back to their repo, we can
hope the issue will also be fixed there timely.
For requests sent by us directly, we include the cookies.
Ref https://github.com/ytdl-org/youtube-dl/issues/28604
