Faster FFT implementation

[carlthome]

I saw that you've thought about this already (#39), @bmcfee, but I hope you'd reconsider, especially for the upcoming ICQT. A key issue with the original MATLAB CQT Toolbox is performance, and librosa could gain traction by providing a high-performing alternative.

Proposal

Preferably all uses (constantq.py, spectrum.py, audio.py) of FFT/IFFT should go through a fourier.py interface, that:

1. Uses cufft if CUDA is available, via scikit-cuda.
2. Then tries to import pyfftw if there's no CUDA.
3. Before finally falling back on scipy.fftpack if there's no libfftw3.

I'm short on time at the moment, but I could turn in a proposal pr in a couple of weeks if no one else has the time.

[bmcfee]

I'm open to the idea, if it can be done cleanly without making the dependency chain unduly complicated.

In the next major revision, I'm trying to pare down the optional/dynamic dependencies (scikits.samplerate and seaborn). I'm not totally keen on the idea of adding more dynamic dependencies back in. Maybe it could all be multiplexed out by a separate library ala audioread?

[carlthome]

I get the ambition to keep librosa clean and portable. I'll ask pyFFTW to add GPU support. pyFFTW/pyFFTW#107 and if that goes through perhaps librosa could exchange scipy.fftpack for pyFFTW (assuming pyFFTW could fall back on numpy or scipy, also, so that libfftw3 remains optional).