Initial Base Classes and base functionality for Neural Network Potentials

[wiederm]

Description

This PR provides an NNP base class, NNPBase, to provide a flexible and extensible foundation for implementing various neural network potentials. It also introduces an implementation of the SchNet architecture for predicting quantum mechanical molecular properties.

Changes:

• Adds Schnet class in schnet.py for implementing the SchNet architecture.
• Introduces a BaseNNP class in models.py as a base class for neural network potentials.
• Includes various utility functions in utils.py like scatter_add, GaussianRBF, Dense, etc.
• Provides test coverage for all the new functionalities.

  Key Files:

• schnet.py
• models.py
• utils.py

[yuanqing-wang]

I personally think that this parent class is a slight overkill.

• We don't need to handle derivatives explicitly, everything will be autograd and should happen outside of the module.
• I don't imagine we would need explicit output collection either.

For the parent class I think it should serve the follow purposes:

• enforce the shape of geometry
• enforce the shape of identity embedding
• allow other universal checks to be implemented

[yuanqing-wang]

@wiederm I'm still a bit confused about the questions I've asked above---could you justify the necessity of those design choices a bit more? Thanks!