One basic library without dependencies, and one with heavy
dependencies like libsamplerate.

This reverts commit c84bae73.  A
configuration option is not necessary, because the PcmConvert
constructor knows already whether integer or floating point is needed.

This code path is 2% faster.

10% speedup.  Strange, huh?

For some reason, this speeds up the library by 2%.

Applying software volume to S16 samples means several bits of
precision are lost; at 25% volume, two bits are lost.  Additionally,
dithering adds some noise.

The problem gets worse when you apply the software volume code twice:
for the software mixer volume, and again for the replay gain.  This
loses some more precision and adds even more dithering noise, which
can become audible (see
https://github.com/MusicPlayerDaemon/MPD/issues/542).

By converting everything to 24 bit, we need to shift only two bits to
the right instead of ten, losing nearly no precision, and dithering is
not needed.  Even if the output device is unable to play S24 directly,
we can convert back to S16 with only one stage of dithering.

Closes https://github.com/MusicPlayerDaemon/MPD/issues/542

Prepare for a new mode which may convert to a different sample format
when applying volume, to reduce dithering.