implement training on per atom energies by modifying loss modules

[SanggyuChong]

Changes implemented here were discussed with @frostedoyster and @Luthaf.

Model training on atomically averaged energies or other extensive targets can be achieved by specifying such peratom_targets, a list of strings, in the training section of the yaml files for the models. Note that this means future models that use the native training routine of metatensor-models should also specify this list in their default yaml files as well.

Then, prior to computing the loss, this list is checked, and for the targets that have been specified, both the model prediction and the reference target values are divided by the number of atoms. This takes place towards the end of the compute_model_loss function.

Resolves #95

---

📚 Documentation preview 📚: https://metatensor-models--105.org.readthedocs.build/en/105/

[frostedoyster]

The changes look good, but we need some documentation. In particular, the current implementation prints the RMSEs per atom of the quantities, while usually we print the RMSE per structure. We should say in the docs that the RMSE of per_atom quantities will also be printed per atom.

We should also check that all per_atom quantities are energies, at least for now

[Luthaf]

@frostedoyster please wait for CI to pass before merging =)

[SanggyuChong]

Thanks everyone!