Flexible property prediction heads

[M-R-Schaefer]

I have added property prediction heads to the EnergyModel.
These can predict per-atom or per-structure quantities with either l0, l1 or symmetric traceless l2 symmetry.

Further, I have added a LatentEwald module which gives us access to long range interactions. It requires a property head which predicts charges.

The datapipeline can now read in arbitrary properties from h5 files. The keywords in the loss function determine which properties are read.

other changes

• slight cleanup of the dtype configs
• moved atomistic model functionality directly into the energy model

[PythonFZ]

Looks good - I left a few questions but no real requests for review. I think this can be merged but I have not tested it locally myself.