Use a separate backend callback to start playback of the device

Use a separate backend callback to start playback of the device

This allows us to properly update the ALCdevice and its resources with the new
parameters before starting playback, instead of expecting the mixer to block
and wait after it has begun.

This also lets us avoid holding the device lock while resetting and starting
the device, which helps prevent lock inversion on some backends (ie, one thread
locking A then B, and another thread locking B then A), ultimately allowing
certain backends to asynchronously update the ALCdevice without risk of lockup.
Capture still has issues here, however.