- Mixes internally in Float32. This simplifies the code immensely by only
having one format to screw with. It also makes life easy for end-user
callbacks. A native Float32 format should be added to SDL, too, so there
isn't unnecessary conversion if we can avoid it (i.e. - a CoreAudio backend).

- "Chunks" are just Sound_Samples...you can lock the mixer to screw with them
(i.e. - seeking in a playing Sample, etc). The mixer adds some opaque state
to Sound_Sample (current play position, how much is decoded, etc), some of
which can be queried and set.

- There is no "stopped" state. You are either in the playing list or you are
not, but state doesn't reset, so removing a sample from the list is more like
pausing it. If you put it back in the playing list without rewinding it, it
starts where it was.

- Fire and forget mixing is easy; flag a sample as "auto free" and it'll
delete itself when it's done playing. No need to set up a callback just to
clean up.

- No channels. You can mix as many samples as you have resources to
accomodate.

- No groups. This can be layered on top of the library if needed. If you
need atomic operations, lock the mixer.

- No music channel. Samples are samples. You can mix a MIDI as a sound effect
if you want, or a WAV file for background music. If you have the horsepower
to decode multiple compressed files at once, go for it.

- You can prebuffer/predecode as much of a sample as you like.

- Every sample mixes with a per-channel gain, plus a master gain that is
global to the mixer.

- Can handle non-power of two resampling.

- post mix hook, sample finished hook. Effects callback?