Use SND_PCM_NONBLOCK, and perform all snd_pcm_writei() calls in the
IOThread.  Use a lockless queue to copy data from the OutputThread to
the IOThread.

This rather major change aims to improve MPD's internal latency.  All
waits are now under MPD's control, instead of blocking inside
libasound2.

As a side effect, an output's filter is now decoupled from the actual
device I/O, which solves a major latency problem with the conversion
filter on slow CPUs and small period buffers.  See:

 https://bugs.musicpd.org/view.php?id=3900