Implementation of register_autograd to recreate autograd graph without recomputation

[agoscinski]

Implements the function register_autograd into the rascaline_torch Calculators that allows to recreate the autograd graph without recomputing the features. This is useful for training over multiple epochs where the gradients can be precomputed and reused for all epochs.

TODOs: I would like to merge EquistoreAutograd into RascalineAutograd, since the backward functions are identical. I would use a check on the nullptr of the tensor_map to determine if it was called from register_autograd or from compute.

---

:books: Documentation preview :books:: https://rascaline--205.org.readthedocs.build/en/205/

[Luthaf]

I would like to merge EquistoreAutograd into RascalineAutograd, since the backward functions are identical.

Yes, that was my idea: refactor the code to have only one custom torch::autograd::Function. I'll finish up #200 first though, this is less urgent IMO.

[Luthaf]

This is now in #200!