Simplify replacement of fftpack by pyfftw

[anntzer]

This patch replaces all instances of from scipy.fftpack import ...
by from scipy import fftpack followed by the use of qualified
names (except in the scipy.fftpack submodule itself). The idea is
to simplify the replacement of scipy.fftpack by the FFTW-backed
pyfftw.interfaces.scipy_fftpack, which provides the same interface.

Currently, acceleration of e.g. scipy.signal.fftconvolve
requires fixing individual functions in scipy.signal.signaltools
(scipy.signal.signaltools.fftn = pyfftw.interfaces.scipy_fftpack.fftn; scipy.signal.signaltools.ifftn = pyfftw.interfaces.scipy_fftpack.ifftn)
because signaltools imports the functions out of fftpack.
With the patch, a simple scipy.fftpack.__dict__.update({name: getattr(pyfftw.interfaces.scipy_fftpack, name) for name in pyfftw.interfaces.scipy_fftpack.__all__}) (which could be wrapped in a
helper in pyfftw's side) should suffice.

[rgommers]

+1 for a few more characters if it makes using pyfftw easier.

My only question is whether a regression test is needed here. No one will remember this a year from now, so a test running on TravisCI only that inspects source code for lines containing import and fftpack will be quite useful in the long run I expect.

[anntzer]

Added tests (regexp based, excluding doctests; the other possibility would have been to use an AST walker).

[rgommers]

Learned about something new (pathlib), nice.

One could argue about wanting to run this on other Python versions and/or decorating it with @dec.slow (test takes 3.5s on my machine). But this does the job, so let's call it good.

Thanks @anntzer!

[anntzer]

This actually takes <0.5s on my machine; perhaps this is due to hard disk access speed?. In fact, I did write a simple AST walker; that took >4s so that's why I came back to the regexp version.

[rgommers]

Hmm, not sure - I do have a fairly modern machine with a solid state drive. But I'm not too bothered by the extra 3 seconds, so won't look into it now. About once a year I get annoyed and start looking at the largest offenders of test suite runtime - this won't be one of them I think.

[hgomersall]

Great, thanks for this. I'll update the docs for pyfftw to reflect the changes.

[rgommers]

We may also want to do the reverse: link to the pyfftw interfaces docs from http://docs.scipy.org/doc/scipy/reference/tutorial/fftpack.html with some explanation. And mention the license differences for background. What do you think?

[hgomersall]

Good idea. I need to push out the new release once I've fixed a small windows bug which I was planning on sorting it tomorrow. I don't think the documentation link would change though.