This is more intuitive because it puts the numerator coefficients
before the denominator coefficients.

With this commit, nothing is wired up with the new code. This will be
coming in the next commits.

Previously this would result in a glitch, however in the interest of
avoiding as many sources of glitching as possible, changing the format
and/or channel count of a biquad filter (and by extension, anything
derived from it, such as low-pass filters) will now result in an error.

This is never actually used internally and saves a bit of memory.

This change reduces memory usage by removing half the number of delay
registers.

This commit also sets up some infrastructure for future work on a fixed
point implementation. Both the floating and fixed point implementations
will run on a 32-bit pipeline for now.

  * ma_convert_pcm_frames()
  * ma_sine_wave_read_pcm_frames()

This changes the the lpfCutoffFrequency config variable from a uint32
to a double. This is required because we simplify the sample rate
fraction which cant result in cases where the cutoff frequency is too
aggressively rounded down due to dividing tiny integers.

This is equivalent to passing in zeroes.

Public issue https://github.com/dr-soft/miniaudio/issues/128