Feat/nls share

[sayeg84]

PR Checklist

• [] Bug fix
• Feature addition/change
• [] Documentation addition/change
• Test addition/change
• [] Black formatting

---

Describe your changes here:

Currently the implementations of the repulsions with cutoffs are naive, in the sense that they compute all the distances and then filter the ones inside the cutoff. It was originally designed like this for consistency with the previous approach, but this is clearly a bad strategy.

This modification allows for sharing the neighbor list between SchNet and the cutoffs by modifying the AtomicData object when it passes through the forward pass and saving the neighborlist in a field called neighbor_list["from_mpnn"], which then is cached by the repulsion and thrown away afterwards.

It should be noted that, given that the repulsion typically excludes some beads that are already restrained by bonds and angles, the neighbor list sharing needs to account for this. Due to this, the neighbor list associated to the repulsion needs now an index_mapping_excluded entry of shape (2,k) with "k" the number of excluded pairs, which will be excluded from the repulsion. This PR also provides a script src/mlcg/scripts/mlcg_add_exclusion_list.py to incorporate this

Empirically, this has increased speed of simulations by a significant factor.

One modification that remains to be done: to generalize this to the other MLIPs