TODO
----

- Make static lookup tables instead of runtime generated ones where aplicable
  to eliminate the small startup latency.

- Use 32-bit multiplications instead of 64-bit ones for less CPU cycles where
  the dynamic range is narrow enough not to affect quality.

- Find out where this small quality difference is lost between the floating
  and fixed point version (need to play a delta of both results to hear it,
  otherwise not noticeable) when lsp is used for certain bit streams.

- Profile CPU usage.

- Check against the official format documentation (when it's available) for
  further possible optimisations.

Contributions welcome as always!