Separate class for core EMLE functionality

[kzinovjev]

The basis for implementing training and error analysis scripts as an integral part of emle-engine. The EMLEBase class provides predictions for valence shell widths, core charges, valence charges, and the 3Nx3N A matrix from the Thole model. It also accepts (padded) batches of coordinates and atomic numbers, instead of a single molecule, which will be necessary for training. The code responsible for calculating the interaction with MM point charges (and therefore not relevant for training) is kept in the EMLE class.

Minor changes:

• There was a subtle bug in how the coefficients for GPR were calculated. This has now been fixed, so the numeric tests will most likely fail. However, I compared the predictions with those from the (JAX-based) training code, and they are identical.
• The alpha_mode='reference' model has been redefined such that the full k is calculated as the product of k_Z and an additional scaling factor. This effectively makes the reference (GPR-based) polarizability model a correction to the species model: if the correction factor is not applied, the species model is recovered. This allows to keep only one default model instead of two separate ones.