Add support for GDK (<https://github.com/microsoft/GDK>)

This will now pick the closest miniaudio-supported format to the source
file's internal format rather than always defaulting to f32.

The issue here is that MA_AT_END is getting returned when nothing has
been read. The rules are that MA_AT_END should only ever be returned
when nothing is read.

This is important for fixed sized processing.

This is the new default. The size of the buffer will be based on the
periodSizeInFrames and periodSizeInMilliseconds config variable. To go
back to the previous system, where the data callback could specify any
number of frames, set the noFixedSizedCallback config variable to true.

This flag is used to tell miniaudio that it should assume the output of
the node is silence. This is useful for special nodes that want to
output to a file and don't want the output of the node to contribute to
the final mix.

Public issue https://github.com/mackron/miniaudio/issues/388

Public issue https://github.com/mackron/miniaudio/issues/386

Public issue https://github.com/mackron/miniaudio/issues/385

Public issue https://github.com/mackron/miniaudio/issues/385

Public issue https://github.com/mackron/miniaudio/issues/385

This was due to context-level mainloop being accessed from different
audio threads at the same time. I've fixed this by giving each device
their own pa_mainloop and pa_context objects.

I considered the idea of having only a single context-level mainloop
and just using a mutex for mutal exclusion, but that would involve a
lock in the audio thread's data loop which I wasn't happy about. Also,
I wasn't sure which thread PulseAudio callbacks would get fired from
since the main loop would be iterated on different audio threads.

Public issue https://github.com/mackron/miniaudio/issues/376