Feature "Moving Window Cross-Spectrum"

[ThomasLecocq]

What does this PR do?

Adding the Moving Window Cross-Spectral method. This method allows comparing two signals in the frequency domain (cross-spectrum, or cepstrum) and returns a table of time delays (offsets) vs time for each window.
See current doc here:
http://msnoise.org/doc/master/core.html#moving-window-cross-spectral-method

PR Checklist

• it is linked to Ambient Seismic Noise in ObsPy #841 and branched from generalized linear_regression #1716

• All tests still pass.

• Any new features or fixed regressions are be covered via new tests.

• Any new or changed features have are fully documented.

• Significant changes have been added to CHANGELOG.txt .

• I should provide an example with some figures.

This PR does look ugly because I actually branched it from #1716 , didn't really know how to do it otherwise as I need the linear_regression in this PR.

[ThomasLecocq]

this fails on Travis while it's OK on appveyor, it is linked to scipy' next_fast_len being available or not. In Appveyor scipy is not fixed, so it uses 0.18.1, while in Travis it seems fixed. The results of my test are different with obspy's next_pow_2 or next_fast_len ... not good.. have to find a way to check for both (and ensure the results are "similar" enough to show the method is stable).

And it's kind of scary that the linear regression is so different for different n submitted to the FFT.

[ThomasLecocq]

@barsch @megies @krischer

The tests fail depending on the scipy version available, because the results of the MWCS are slightly different when using nextpow2 or next_fast_len ? Should this function be defined with nextpow2 by default but allow for some sort of callback in case one wants to use next_fast_len (scipy >= 0.18) ?

[krischer]

Can you up the test tolerance to have it test similar enough on different scipy versions?