Raise error when running inference on elements without fitted element references

[lbluque]

Summary

Add NaN-based detection in ElementReferences.compute_references that raises a clear ValueError when input data contains elements with no fitted references, indicating the model was not trained on those elements.
Change fit_linear_references to initialize unfitted element coefficients to NaN (instead of zero) so untrained elements are distinguishable from elements with a legitimately fitted zero reference.
Replace the previously skipped test_element_references.py with 6 new working tests covering both error and success paths.

Motivation

Previously, running inference on a system containing elements the model was never trained on would either silently produce incorrect predictions (the reference contribution for that element would be zero, which is wrong) or crash with an opaque CUDA/index error if the atomic number exceeded the embedding table size. This change ensures users get a clear, actionable error message identifying exactly which atomic numbers are unsupported.

TODO

• Insert NaN values to element reference files on HF.