ESMFold2: a ligand given as several CCD codes is returned as one residue #419

A ligand given as several CCD codes comes back from fold() as one residue named after its first component, holding the atoms of all of them:

LigandInput(id="B", ccd=["NAG", "BMA"])
# complex.sequence: [..., "NAG"]           # both components' atoms under one name
# expected:         [..., "NAG", "BMA"]

The tokenizer already keeps each component's residue_index; build_molecular_complex_from_features in esm/models/esmfold2/output.py collapsed every non-polymer chain into one residue. A side effect is that to_mmcif() keys entities on residue names, so ["NAG", "NAG"] and ["NAG", "BMA"] are written as the same entity.

PR #416 groups non-polymer tokens by residue_index, the way atom-tokenized polymer residues such as MSE already are. SMILES ligands, single-CCD ligands and polymers come back as before. Happy to change it if you would handle this differently.