simplify model prediction routine on the python side

[felixmusil]

The only model at the moment (the KRR class) has only one predict function that depending on the inputs return model values, derivatives with respect to atoms positions (and stress). This is a bit messy and to simplify it, it should be split into several dedicated functions. Also we might want to have a simple KRR that handles only models without gradients (so typical QM properties like chemical shieldings or band gap) and a GAP class that focus on energy models, aka inter atomic potentials ?