implementation of training on the per atom target quantities

[SanggyuChong]

Hello,

This PR is a first attempt at the implementation of training the models (SOAP-BPNN and alchemical) on the per atom targets for extensive quantities (e.g. energy) rather than the raw extensive values. Usefulness of this, apart from personal preferences in model training, is that it allows one to retain consistency between models in metatensor-models and default training behavior of MACE, NequIP, and other packages.

In the proposed solution, changes are summarized as follows:

• peratom_targets, a list of strings that contain targets that should be trained on the per atom values, is defined and accepted as an input to comput_model_loss function of compute_loss.py. The list can be supplied in the input yaml file.
• in the compute_model_loss function, model predictions and the target values are divided by the number of atoms before it is passed to loss then MSELoss.

I understand this may not be the most optimal solution to this. All suggestions are welcome in refactoring the feature.

Resolves #95

---

📚 Documentation preview 📚: https://metatensor-models--101.org.readthedocs.build/en/101/

[SanggyuChong]

closing this PR as @frostedoyster and I have decided to modify a completely different part of the software stack.

[SanggyuChong]

After another discussion, we decided to keep the original approach, which has been incorporated into a new PR under #105.