Adds a floating-point API to Opus

The high-level Opus encoder and decoder can now be compiled as
either fixed or float. Also, we now use the stack_alloc.h macros
in the top-level Opus code.