bugfix for stress loss

[JPDarby]

Removed division of stress by n_atoms in loss to resolve #437

[bernstei]

I noticed that the train.py error report says stress_per_atom, which makes me worry - stress is intensive, it shouldn't be "per atom" the way I usually use that phrase. As I poke around the code I see delta_stress and delta_stress_per_atom, e.g.

mace/mace/tools/train.py

Line 422 in dee204f

self.delta_stress.append(batch.stress - output["stress"])

Can someone (@ilyes319 ?) confirm that this kind of bug isn't present elsewhere in the code?

I'm hoping that those variables mean "total stress" and "atomic stress", rather than per_atom meaning divided by n_atoms. However, even if it is, depending on what exactly the atomic stress is (i.e. dE_site / dVol_total, or dE_site / dVol_site (= V_total / n_atoms?), it may or may not be what people are expecting.