Refactor CompositionModel

[jwa7]

The current implementation of CompositionModel is becoming problematic and difficult to extend with the widening use cases of targets. As such, a refactor of CompositionModel is needed to accommodate these and maintain/improve modularity.

As discussed, the idea is to have a single CompositionModel class that dispatches the fitting procedure to different functions depending on the target type - such as global, per-atom, or per-pair quantities. This should to be a standalone module, but integrate with existing models in metatrain. The 'weights' of the model can be stored in a TensorMap, but properly registered as a buffer such that they are frozen during training but available in an exported model.

Tagging those involved @frostedoyster @Luthaf @ppegolo @SanggyuChong. Please comment if I've missed something!