Fix PDB parsing for multi-chain polymers

[ryanhulke]

fixes #236.

PDB files can declare one polymer entity with multiple chains, e.g. COMPND ... CHAIN: A, B. The PDB parser was assigning label_seq only for entity.subchains[0], leaving later subchains without label_seq_id after Gemmi converts the PDB to mmCIF.

this caused residues in the following subchains to be parsed as is_present=False and then dropped during mmCIF serialization, which could create token/design-mask length mismatches downstream.

Changes

• assign label_seq for every polymer subchain in a multi-chain entity
• preserve the existing fallback for generated PDB's without full_sequence/SEQRES by inferring the sequence from the first non-empty subchain

Testing

• reproduced with PDB 3K4P, which declares CHAIN: A, B
• confirmed parse_pdb(..., use_assembly=False) preserves both chains,
  • A: 444 residues, 438 present
  • B: 444 residues, 438 present
• confirmed to_mmcif(...) round-trip preserves both chains with the same counts
• ran:
  • python -m py_compile src/boltzgen/data/parse/pdb_parser.py
  • python -m pytest tests/test_residue_constraints.py
  • python -m pytest tests/test_inverse_fold_constraint_masks.py