I confirmed that the currently used real FFT code in RealFFTf.cpp is faster
than the old one with a quick benchmark that calls PowerSpectrum() on 4-minute
audio file, with different sizes of computation windows:
Window_size: 256 method: new FFT time_s: 0.393
Window_size: 256 method: old FFT time_s: 1.065
Window_size: 1024 method: new FFT time_s: 0.38
Window_size: 1024 method: old FFT time_s: 0.958
Window_size: 4096 method: new FFT time_s: 0.413
Window_size: 4096 method: old FFT time_s: 1.084
Window_size: 16384 method: new FFT time_s: 0.518
Window_size: 16384 method: old FFT time_s: 1.338
Window_size: 65536 method: new FFT time_s: 0.655
Window_size: 65536 method: old FFT time_s: 1.524
Window_size: 262144 method: new FFT time_s: 0.735
Window_size: 262144 method: old FFT time_s: 1.873
Create WaveTrackCache as a utility class but don't use it anywhere yet.
The possible minor performance problem with effects is fixed by changes
in WaveTrack::GetBestBlockSize().
