kM is used in various computations, kMspec is the one relevant for
spectra.
And they are different for the smooth filter case...

I'm not using the ideal estimate of the integral length since I'm doing
the integral with the 1D spectrum, but I think this should be close
enough.

in practice nothing changes, but the spectra will be shorter and the kM
eta calculation will be "correct".

it's now based on the last git commit

rFFTW_interpolator->descriptor is not allocated in the constructor,
therefore it should not be deallocated in the destructor

it's not more efficient.

I don't really like it this way, but otherwise it won't compile when
needed.

somewhat.
Python 3 insists on generating UTF-8 datasets with h5py, unlike python
2. I guess this commit deals with the problem.

io fails at the moment