The unbounded apad filter made ffmpeg pad the 3-second test clip
endlessly (a 22.9GB WAV) until the disk filled, which was the real
cause of the voice separation failures. Trim to the probed duration
and drop apad.
The private FunASR/HTDemucs audio service was already live and used by
the smart edit pipeline, but the canvas video toolbar still claimed
voice/background separation was unavailable. Add a dedicated
/api/canvas/audio-separate route (download, ffmpeg stem extraction,
remote separation, storage upload) and route the toolbar's voice and
background modes through it, replacing the stale not-connected notice.
