This will make it easier to integrate into the enhanced branch, and
also improved audio quality slightly (samples are mixed as 16-bit
instead of 8-bit).
The mixer itself is integer-only, but the millibel stuff is still
float-based (I don't know if it's worth adding a 10001-value-long
lookup table to replace it).
Anyway, according to a quick test, this new mixer is significantly
faster than the old floating-point one - rarely going above 100
clock() ticks per callback, with the old one almost always running
above that.
For now, only the SDL2 backend supports it - I'll add the others in
upcoming commits.
This will no-doubt cause problems with the enhanced (lite) branch,
where clownaudio always outputs float samples.
