Instead of passing individual buffers to audio_output_all_play(), pass
music_chunk objects.  Append all those chunks asynchronously to a
music_pipe instance.  All output threads may then read chunks from
this pipe.  This reduces MPD's internal latency by an order of
magnitude.