Feat/sparsified soap

[felixmusil]

This PR allow for the sparsification of the powerspectrum.
Changes on the C++ side:

• loops over (n1, n2, l) in the computation of powerspectrum are now over a compound index that contains all the information for indexing the powerspectrum and the spherical coefficients.
• by default this compound index is initialized to replicate the former loops over (n1, n2, l)
• if 'coefficient_subselection' with form {'a': [...], 'b': [...], 'n1': [...], 'n2': [...], 'l': [...]} is in the hypers then only the coefficients of the powerspectrum with species (a,b) and (n1,n2,l) will be computed. The resulting powerspectrum has one dummy key (0,0) which contains all the selected coefficients.
• all of these 'features' are tested in test_calculator.cc using a slightly different take on the test organization. Tell me what you think please.

Changes on the python side:

• the powerspectrum know how to create a mapping between dense linear index to the corresponding (a,b,n1,n2,l)
• a FPS wrapper class that will be used to showcase the sparsification of the features in a notebook

[max-veit]

Part the Second, in which the C++ code is reviewed.

Ok, I can follow what's going on for the most part, although some features need better documentation (see comments). It seems that the sparsification has the nice side effect of reducing the nesting depth in many loops, which definitely aids understanding. New test structure looks good and is much easier to understand.

[max-veit]

Um, how about changing the Python formatter in a different branch and PR? There's enough changes mixed up into this one already.

[felixmusil]

Sure that's how it should be done.

[max-veit]

Then why does the latest commit add black.cmake? It's hidden in a commit with the undescriptive name comments, so it'll be hard to track where it got added from the log.

[felixmusil]

right got added by mistake I guess.

[max-veit]

now OK to merge by me