Enable building against MKL-enabled Anaconda

[mjuric]

This is fundamentally an awful hack, but it makes it possible to build conda packages that work with MKL versions of numpy, scipy, etc. We:

• Minimally patch LSST's version of the fftw3 library to a) prefix every symbol with lsst_ (but #define it to the usual name) and b) only build the static library (so we don't accidentally have non-LSST code linking against our FrankenFFTW).
• Patch psfex (the ./configure script) to detect the hacked up library. Other code using libfftw (e.g., tmv or ndarray detects it just fine).

Note: I worry that our issues with MKL could be indicative of a real problem on our end. Our code doesn't crash, it just returns (very) incorrect results.

It may be that we're just not checking the return codes from FFTW3 calls and detecting when they fail (MKL doesn't implement everything FFTW3 proper does), but it's also possible we're (incorrectly) assuming something about FFTW3's memory management that just happens to work with the reference implementation but not MKL wrappers.

This could be worth understanding; the number of places where we're using FFTW isn't that large (I think it's only for sinc interpolation), so it may not be very difficult.

[mjuric]

A better way to handle this has been put forward as a part of #73.