Add 1xnr model (new format)

[IgnacioJPickering]

• Add ani 1xr
• Fix issues with 1xnr model
  This PR superseedes Add ani-1xnr model #653
  @JMorado

This PR adds support for ANI-1xnr.

I wanted to double check what was the level of theory this model was trained on (docstring says BLYP-TZV2P) since I'm getting "strange" energy results. The absolute energy values are significantly different from 1x and 2x. Is this normal?

For example for the following molecule:

        species = torch.tensor(
            [[6, 8, 7, 1, 1, 1, 1, 1]], dtype=torch.long
        )
        coords = torch.tensor(
            [
                [
                    [0.8715798227, 0.1619958512, 0.0322465836],
                    [0.0824525507, -0.6745183984, -0.7728562897],
                    [-1.2366803437, -0.3917257278, -0.6407828312],
                    [1.3846786207, -0.5081762833, 0.7522303382],
                    [0.2593533005, 0.9062281629, 0.5764446979],
                    [1.6258014512, 0.7033308948, -0.5701975001],
                    [-1.6685247073, -0.8348361119, 0.1963280892],
                    [-1.3186606949, 0.6377016126, -0.6093561983],
                ]
            ],
            dtype=torch.float32,
        )

The energy is around -34 Ha, but it is around -170 Ha for 1x and 2x. The atomic self energies are also significantly different.

If this looks good to you, you can add a commit so the PR is shared, and I will merge (I have given you write-access to the repo, so you should be able to do this, no problem).

Note that the actual model used in the repo now lives on https://huggingface.co/roitberg-group/ani1xnr since the neurochem format is a pain, so I converted all models to pytorch state dicts

[IgnacioJPickering]

@JMorado I've cherry picked one of your commits so that you show up as co-author of the PR. I'm still not sure about the model however, the forces seem logical, but the energies are very strange.

I thought that maybe since the model was trained with PBC, with VASP? or something like that, maybe you used some approximation in the basis set/level of theory that shifts all the energies.

I've calculated the same molecule with orca with BLYP TZVPP and I get a result similar to ANI2x / ANI1x