test(rf3): add MSA input tests

[woodsh17]

Summary

Adds tests for the MSA (multiple sequence alignment) input that rf3 accepts at
inference. There was no coverage of the MSA input path and no example that
documents the expected .a3m format.

• Layer 1 — a3m format-contract unit tests (checkpoint-free).
  models/rf3/tests/test_msa_format.py pins the ColabFold-style .a3m format
  and how atomworks.parse_a3m reads it, using a small self-documenting
  fixture (query + 3 homologs) that exercises every parsing rule:

  • MSA shape (N, L) with the query as row 0 and uniform column width;
  • lowercase insertions are stripped and counted; - gaps are kept as columns;
  • TaxID= taxonomy IDs are parsed from headers (the key that pairs homologs
    across chains for multimers), including the query sentinel and the
    empty-header case;
  • a query-only alignment parses to depth 1 rather than crashing.

• Layer 2 — end-to-end fold integration tests (checkpoint-gated).
  models/rf3/tests/integration/test_msa_fold.py feeds pre-computed MSAs
  through rf3 fold across chain counts (monomer / homodimer / heteromer) and
  both input modes (JSON msa_path and CIF _msa_paths_by_chain_id):

  • monomer via JSON msa_path;
  • monomer via CIF _msa_paths_by_chain_id;
  • homodimer (two identical chains sharing one a3m), asserts iptm > 0;
  • heteromer (two distinct chains, each with its own a3m, sharing a TaxID to
    form a genuine paired MSA), asserts iptm > 0;
  • the raise_if_missing_msa_for_protein_of_length_n guard: errors when a
    protein above the length threshold has no MSA, and stays quiet when present.

  A msa_fold_dir fixture batches the four folds into one model load;
  msa_present_flag_dir and an assert_valid_plddt helper are added alongside.

Test data

All inputs use short synthetic sequences (~12 residues) with hand-built a3m
files: these tests exercise MSA format and plumbing, not structure quality, so
a real complex — which is far slower on CPU — is unnecessary. The CIF-mode
input was generated by folding the synthetic monomer once and attaching the
_msa_paths_by_chain_id header (the same construction as the existing
*_from_file.cif inputs).