DSound: use DSound notifications to produce sound.

[magumagu]

Pretty straightforward; IDirectSoundNotify lets you register for
notifications after a certain amount of sound has played, so use that
instead of depending on Update() notifications from the CPU thread.

Also, while I'm here, reduce the buffer size by a factor of 4; this seems
to reduce the latency, although the difference is sort of subtle.

[shuffle2]

The IDirectSoundNotify concept LGTM.
However, BUFSIZE(0x2000) is not evenly divisible by 3, will that be a problem, or is the last interval being longer OK?
I'm not sure how lower BUFSIZE affects people on faster/slower hardware; are we protected from under/overflow?

[magumagu]

It doesn't really matter whether the intervals are precisely the same size; we just fill the buffer every time the notification triggers.

The DSound buffer is circular, and we always just fill the buffer, so overflow is fundamentally impossible. If we don't fill the buffer in time, the sound will get interrupted... but that's unlikely to happen very often even with a weak CPU. The OS scheduler should give the sound thread enough CPU time because it's sleeping most of its time. (Dolphin's Mixer class has its own independent buffer which can overflow and underflow... but it handles all of that internally, so it isn't really relevant here.)