This improves the performance of the analysis (by 2x!),
by offloading non-`libebur128`-computations (i.e. decode-convert)
to a separate thread, thus reducing walltime.
The most juicy bit!
This is based on Song Fingerprint Analysis,
but here we must know the actual song, and not just the file.
The library supports only interleaved S16/S32/F32/F64,
so we must be sure we insert `audioconvert` into pipeline.
One point of contention here for me, is whether we should
feed the frames to the library the moment we get them
in `NewBufferCallback`, or collect them in a buffer
and pass them all at once. I've gone with the former,
because it seems like that is not the worst choice:
https://github.com/strawberrymusicplayer/strawberry/pull/1216#issuecomment-1610075876
In principle, the analysis *could* fail,
so we want to handle that gracefully.
