This involved adding kfft_single.[ch] that redefines kiss_fft a second time
with a different prefix. All this is still a bit of a mess now. The mask
had to be converted to 16-bit input, but we're still using floats to apply it.

handling of the left side slope of the spreading function (right-side coef
was used)

itself

cos() and sin() functions.

speexdsp library.

copies of the FFT.

window size when parts of the window are zero.

anyway (check-point for pitch improvements to follow)

zeros and ones (nor even the second half).

(sort of)

encoders and decoders.

them saves another 1kB for fixed-point (2kB for float or full precision)