Implement saving and loading of pyFFTW wisdom for substantial speedup

[smallnamespace]

As an example, running forward NSGT on fugue.wave takes 1s compared to about 180s initially, and nsgt_test.py takes 8s instead of 30s on my machine, because we essentially skip over all the planning.

[sevagh]

I'd like to try to get this working on my computer - once it does, I'll give additional thumbs up for this to get merged.