libsamplerate 0.1.2 didn't have the 32 bit <-> float conversion
routines.  Emulate them in case they aren't supported.

Similar to pcm_resample_16(), implement pcm_resample_24().  The 24 bit
implementation is very similar, but it uses src_int_to_float_array()
instead of src_short_to_float_array() before sending data to
libsamplerate.

A future patch will implement a 24 bit resampler.  To unify code, move
code which can be shared to a separate function.

Added assertions which ensure that there are no partial samples in the
source buffer.

Use sizeof(sample) instead of hard-coding "2".  Although we're in 16
bit right now, this will make code sharing easier when we support
other sample sizes.

Separate the resampling code from the rest of pcm_utils.c.  Create two
sub-libraries: pcm_resample_libsamplerate.c and
pcm_resample_fallback.c.