Update variants

[PicoCentauri]

This PR adds a couple of updates to the variants:

• Allow no default unit present: i.e. now allowed are outputs like "energy/A", "energy/B" without "energy" being present.
• Add a backwards compatible description attribute to ModelOutputs.
• Added pick_variant function that will pick a variant based on a name (like "energy"), a ModelCapabilities object and an optional desired_variant. The function will try to pick the user choice and if not given either pick the non-variant or error giving the provided variances. I am not happy with the parameter naming and order. We can discuss it

TODO

• Update CXX tests
• Use pick_variant in ase calculator
• Update Python tests
• Update docs for variants

Reviewer checklist

• CHANGELOG updated with public API or any other important changes?

📚 Download documentation for this pull-request

[PicoCentauri]

The pick_output will raise an error here if the uncertainty_threshold is a number (which is the default) but the model does not have an energy_uncertainty in its outputs. I think this is what one actually wants. If a user desires an uncertainty warning but the model can't do it - we should error.

But this change kinda leads to that one now have to set the uncertainty_threshold to None for all models without an uncertainty output. Before we silently ignored this. We can also change the default of uncertainty_threshold from 0.1 to None.

[Luthaf]

So I think the behavior should be

• by default, if the model has an energy_uncertainty output, we use it with a threshold of 0.1 eV/atom without the user having to do anything.
• if the users specifies a variant for energy_uncertainty but does not touch the threshold, we use the same 0.1 eV/atom threshold
• the user can explicitly opt-out of the warning by setting the threshold to None.

---

In the code, we could try to check if there is an energy_uncertainty output before calling pick_output to resolve it.

[PicoCentauri]

Okay we can do a check if energy_uncertainty or a variant is in any of the output keys.

[Luthaf]

I would only check if one variant is present, the one that would be used by pick_outout.

[Luthaf]

So I think the behavior should be

• by default, if the model has an energy_uncertainty output, we use it with a threshold of 0.1 eV/atom without the user having to do anything.
• if the users specifies a variant for energy_uncertainty but does not touch the threshold, we use the same 0.1 eV/atom threshold
• the user can explicitly opt-out of the warning by setting the threshold to None.

---

In the code, we could try to check if there is an energy_uncertainty output before calling pick_output to resolve it.

[Luthaf]

I would not call this units, variants was fine though

[Luthaf]

we require torch >=2.1, so this check is not needed