Training silently ignores pbc/cell keys missing from the H5 file

[Jamba15]

When data.loading_options.hdf_key_mapping declares pbc and cell keys but the underlying H5 datasets don't actually exist, training proceeds without error.

If pbc/cell are declared in hdf_key_mapping, the loader should validate that the corresponding datasets exist in the H5 before training starts.

[sayeg84]

Thanks for raising the issue.

This is indeed a problem, due to the current design of the dataloader, which is too permissive and work with both PBC and non PBC data.

It would make sense to include this into a complete refactor of the dataloaders, which currently also need to be modified given that we have the noise dataloader on feat/noise_up.

Regards, Aldo.