ESMFold2: fold() cannot take precomputed LM hidden states #420

EsmFold2Model.forward accepts lm_hidden_states, but ESMFold2InputBuilder.fold() has a fixed keyword list and never forwards them. To reuse one ESMC pass across several folds (seeds, sampler settings, ensembles), or to fold with a trunk whose ESMC lives on another device or process, callers have to drop to prepare_input + forward + decode by hand.

A related gap: when the model has no ESMC attached (model.esmc is None, e.g. from_pretrained(..., load_esmc=False)) and no states are given, fold() skips the LM pathway without any signal and returns the prediction of a different computation than the trained model.

PR #415 adds lm_hidden_states to fold() with input validation, makes fold() raise when it has neither an ESMC nor states, and adds a public compute_lm_hidden_states() for computing them in a separate step. The hard error is a behaviour change; an explicit opt-in is easy to add if you would prefer that. I would be glad to reshape the API to whatever fits best.