Public issue https://github.com/mackron/miniaudio/issues/349

This no longer stores per-channel data within the object. Instead it is
now allocated on the heap.

Public issue https://github.com/mackron/miniaudio/issues/348

Public issue https://github.com/mackron/miniaudio/issues/347

It was only disabling denormals when "noDisableDenormals" was set, which
is the opposite of what was intended.
Signed-off-by: Steven Noonan <steven@uplinklabs.net>

We can use _controlfp_s instead of _mm_getcsr/_mm_setcsr for
architecture-agnostic control of denormal behavior on Windows.
Signed-off-by: Steven Noonan <steven@uplinklabs.net>

Signed-off-by: Steven Noonan <steven@uplinklabs.net>

Signed-off-by: Steven Noonan <steven@uplinklabs.net>

This catches the case when backends do not return MA_AT_END when no
data is read.

This updates ma_encoder_write_pcm_frames() to be consistent with
reading, where a result code is returned, and the number of frames
written is returned via an output parameter.

With this change, an error is now returned when the requested frame
count is zero. Two reasons for this:

  1) It usually means there's an error in the higher level logic if
     something is trying to read zero frames.

  2) When no frames are read, MA_AT_END should be returned. However, if
     the input frame count is also zero, it creates ambiguity as to
     whether or not the data source is truly at the end.