ZBL follow-up

[frostedoyster]

#335 leaves a few problems open for ZBL:

• units
• should the composition model be fit before or after removing ZBL? At the moment, it's "before" for the native models and "after" for PET
• the current version breaks distributed training by working on CPU (DistributedDataParallel doesn't like it)

[frostedoyster]

A PR will be opened soon. Regarding the 3 points above:

• units: it's too early to have an internal unit system in metatrain. For now, users that want ZBL will have to work with eV and A
• the composition model will always be fitted after ZBL is removed
• will be fixed

[Luthaf]

units: it's too early to have an internal unit system in metatrain. For now, users that want ZBL will have to work with eV and A

Do we send an error if the users has his dataset using different units?

Also, I don't think this would require having a full internal units systems in metatrain. Instead we could use metatensor.torch.atomistic.unit_conversion_factor("energy", "eV", "user-unit") when initializing the ZBL module to convert between the data stored in the code and what the user wants to use.