Better FFT interpolation for singular integration

[f0uriest]

FFT interpolation can't handle non-symmetric grids. We will probably always need the source grid to be symmetric to capture the integral correctly, but in theory the evaluation grid could be symmetric if we get the interpolation to work correctly.

Originally posted by @ddudt in #725 (comment)

[dpanici]

can't handle symmetric grids right?

[unalmis]

I think the better solution is

1. Improve FFT interpolation (1/2) f0uriest/interpax#116
2. Improve FFT interpolation (2/2) f0uriest/interpax#117

One result of those pull requests is that you can set source_grid(M=40, N=40) and eval_grid(M=20, N=20) and still have a completely lossless interpolation.