On OSX, when the audio output device changes, the OS will call the audio
callbacks in weird patterns, if at all, during a period of ~1s. If
real-time SourceMediaStreams are present in the MediaStreamGraph, this means
buffering will occur, and the overall latency between the MediaStreamGraph
insertion time, and the actual output time will grow.
To fix this, we detect when the output device changes, and we switch temporarily
to a SystemClockDriver, that will pull from the SourceMediaStream, and simply
discard all input data. Then, when we get audio callbacks called reliably
(basically, when OSX is done switching to the other output), we switch back to
the previous AudioCallbackDriver.
We keep the previous AudioCallbackDriver alive using a self-reference. If an
AudioCallbackDriver has a self-reference, that means it's in a state when a
device is switching, so it's not linked to an MSG per se.
