Account for precomputed pure element reference configuration energies #17

Can we make it an issue so it does not block the release of v0.5.0? I think we can do it, but I didn't want to get it rushed and have to break anything in v0.6.0. Probably it is actually better if we implement this directly to v0.6.0 considering it has to orchestrate more models/constraints.

Another cheap option is to keep only the DFT values of these references, which already exist in the matbench_discovery reference we use. I just didn't add them to this fixture bc we were not using this data, but it may serve us to prevent the model drifts you mentioned.

I think we should:

1. Have a list of precomputed pure element reference configuration energies that correspond to every model we have defined in crystALL
2. Have the above done by a GitHub Action specific to a model, every time we (a) modify the references or (b) release a new model
3. Kinda following on point 2; allow users to recompute them locally to allow for new model versions and new reference versions if needed

Also:

• I think that ensuring users the references are fixed for each crystALL version we ship is our top priority - computing them on the fly wouldn't be too bad performance-wise but if they change silently that would be very bad
• Round them to account for numerical precision across machines.

Originally posted by @amkrajewski in #10 (comment)