Weights are now converted ahead of time by convert_esmc_checkpoint.py / convert_esmfold2_checkpoint.py instead of by a runtime conversion mapping: upstream module names renamed, packed q/k/v and SwiGLU gate/up split, TransformerEngine _extra_state dropped, and config keys renamed off d_model/n_heads/n_layers. Verified bit-exact against the previous mapping-based load.
