This is nearly 4 times faster than the "portable" algorithm.

The previous commit eliminated the need for that.

.. instead of reusing FloatToInteger<S24> and converting from S24 to
S32 in-place.

Refactor the conversion functions to classes and pass an instance to
the new function AllocateConvert().

The PcmResampler interface is implemented by the two classes
FallbackPcmResampler and LibsampleratePcmResampler.  This prepares for
adding more resampler libraries.

Copy the according file from another project (i.e. XCSoar).  This will
allow copying more code more easily.

DSD-over-USB should not be a MPD core format, because it is not a
"natural" format; it is just a temnporary over-the-wire format.  This
format has been implemented in pcm_export, and does not need to be
supported by pcm_convert.

For simplicity, the MPD core should not have to deal with packing.  It
is rarely used, and those plugins that need it should use the
pcm_export library instead.

Eliminate support for reverse endian samples from the MPD core.  This
moves a lot of complexity to the plugins that really need it (only
ALSA and CDIO currently).

This format is unused since the DSDIFF decoder plugin now reverses the
bit order.

Basic support for Direct Stream Digital.  No conversion yet, and no
decoder/output plugin support.

Support for conversion from float to 16, 24 and 32 bit integer
samples.

Reduce the mess with local variables inside switch{}.

Allow more compiler optimizations.

This is easier and more efficient to loop on, because only two
variables get modified (src and dest).