As of this commit, ma_engine_sound_set_fade_out() is not fully
implemented pending the writing of some additional infrastructure.

This is useful for having a separate fade in and fade out for a sound
which is what we require for the ma_engine API.

This adds support for having a sound fade in when it is started and
fade out when it is stopped.

This commit does not yet include support for fading out when the sound
approaches the end - it only fades out when explicitly stopped with
ma_sound_stop().

The fade time is set in milliseconds.

This commit includes a new effect called ma_fader, but it currently
only supports f32 formats. Support for other formats will be added in
the future.

The ma_async_notification object is used for notifying the application
that an asynchronous operation has completed.

Custom notifications can be implemented by implementing the callback in
ma_async_notification_callbacks. There is currently only a single
callback called onSignal which is fired when the operation completes. A
helper notification which wraps around an ma_event object called
ma_async_notification_event is implemented which you can use as an
example for building your own notifications.

  * ma_audio_buffer_get_available_frames()
  * ma_decoder_get_available_frames()

  * The data buffers and data streams are now first class data sources.
  * The ma_resource_manager_data_source object is now just a simple
    wrapper around ma_resource_manager_data_buffer and
    ma_resource_manager_data_stream.
  * Unnecessary pResourceManager parameters have been removed.
  * The part of the data buffer that's added to the BST has been split
    out from the main data buffer object so that the main object can be
    owned by the caller.
  * Add ma_resource_manager_data_source_get_available_frames() which is
    used to retrieve the number of frames that can be read at the time
    of calling. This is useful in asynchronous scenarios.

  * A crash when the pFramesRead parameter is null
  * Looping not working due to MA_AT_END now being returned.

Public issue https://github.com/dr-soft/miniaudio/issues/185

Public issue https://github.com/dr-soft/miniaudio/issues/184

This should fix the case where multiple context's are initialized and
uninitialized at the same time across multiple threads.

Public issues:

  * https://github.com/dr-soft/miniaudio/issues/163
  * https://github.com/dr-soft/miniaudio/issues/183

This commit changes synchronous decoding so that the calling thread is
the one which performs the decoding. Previously, decoding was done on
the job threads which was then waited on by an event on the calling
thread. The rationale for this design was to keep decoding on a single
code path, however this creates a problem for programs that would
prefer not to have any asynchronous job threads. In this case, these
synchronously decoded sounds would never get decoded because there
would not be any threads available to actually perform the decoding.

This commit enables the resource manager to be able to be used without
a job thread so long as asynchronous decoding and streaming are not
used. This scenario could be useful for programs that want to pre-load
all of their sounds at load time and save some system resources by not
incurring the overhead of an additional unnecessary thread.