Add Lennard-Jones implementation and training of dispersion coefficients

[JMorado]

This PR adds the Lennard-Jones EMLE, enabling it to function as a fully coupled scheme in which both electrostatics and dispersion are handled by emle-engine. There are two LJ modes:

• flexible: the LJ parameters for the ML region are determined on-the-fly and are geometry-dependent
• fixed: the LJ parameters for the ML region are fixed, and can either be provided by the user or determined from a given geometry (e.g., an optimised structure)

The main change compared to how the code worked before is that the forward pass of EMLE now always returns the LJ energy as the last output, which will be zero if no LJ mode is specified (the default). Everything else remains the same, meaning that old scripts should still work, as all LJ settings are optional.

I've also added training routines for the required parameters. The learned property is the dispersion coefficient (e.g., C6), so one value per reference AEV is added to the model and learned by direct fitting to the targets.

I'm still expecting some changes in the future as this is being tested, so I'll leave this as a draft for now. Compatibility with Sander still needs to be added.