GitLab

This was very confusing for a very long time, and apparently based
on assumptions about SDR#. In the end, our phase was always defined the
wrong way around (clockwise, mathematically negative). This leads to a whole
lot of weird bugs and code that should be different.
Major results:
 - downmixing requires negative frequency
 - FFT is correctly performed with fftw_forward
 - Sample format conversion is faster (eliminated the swap)