Added cg_utils to Rascaline

[ecignoni]

---

📚 Documentation preview 📚: https://rascaline--185.org.readthedocs.build/en/185/

[Luthaf]

This is missing a test to ensure the code actually does what we think it do!

[Luthaf]

Linter is not happy, you can run it locally with tox -e lint and fix most of the issues with tox -e format

[agoscinski]

Writing down TODOs because I will probably not touch it for a week

TODOs for this PR:

• decouple(couple(X)) propto X
• compute SOAP and compare with spherical expansion + CG utils
• regression test
• refactor a bit
• remove nonzero mask
• add code documentation from Edoardo's code cgreal.zip
• sync with documtenation in PR Added doc explanation on rotation equivariant features #184

TODOs in following PRs:

• add equistore block support https://github.com/lab-cosmo/equistore-examples/blob/restructured-pytorch-hamiltonian/torch_hamiltonian_utils/torch_ham.py
• add torch support
• think about support of n-centers representations with this (imn, ijm'n' -> ijµnn' is one of them, but there are a lot of variations)

[agoscinski]

So there is a more documented version in librascal https://github.com/lab-cosmo/librascal/blob/master/bindings/rascal/utils/cg_utils.py Also with two very good notebook examples. nice_demo and equivariant_demo. I will try to merge these two into the current version.

I still need more time to understand the code, its quite complicated. I think we can completely get rid of this combinations_string by enforcing an order of the dimension in the array. The flexibility by the argument makes it harder to understand and to use. From the examples I cannot really gather a real advantage for this flexibility.

[Luthaf]

Closing in favor of #237