The rationale for this is that it's to unreliable and inconsistent
across each backend and not worth the added maintenance cost.

These changes add a mal_context parameter and moves the output device
to the last parameter for consistency with mal_context_init().

Rationale:
1) It allows the callbacks to be set at initialization time which feels
   a bit more intuitive to me.
2) It avoids the need to call mal_device_set_send_callback(), etc.
3) It's a bit more consistent with the onLog callback. Previously,
   onLog would be passed to mal_device_init(), whereas onSend, etc were
   set with mal_device_set_send_callback(), etc. which feels needlessly
   inconsistent.

Rationale:
1) These formats just aren't common enough to justify the added
   maintenance cost.
2) It makes software format conversion a lot simpler since it reduces
   the number of permutations.

Previously, there was only a single default buffer size and period
count. This forces the library to use a lowest common denominator for
every backend which forces higher quality backends to suffer because
of the lesser quality ones.

This change adds basic infrastructure for future work to tweak these
properties for each individual backend to try and get the best out
them.

Rationale:
1) The number of parameters is just getting too much.
2) It makes it a bit easier to add new configuration properties in the
   future. In particular, there's a chance there will be support added
   for backend-specific properties.