Turn into package and improve chain handling.

[colbyford]

ipSAE Improvements

Installable library

• src/ipsae/ package (utils, parsers, confidence, scoring, outputs, api, cli) with pyproject.toml. This installs via pip install ., giving the ipsae console command and python -m ipsae. Root ipsae.py remains a fully backwards-compatible wrapper.
• Programmatic API: score_interactions() returning structured results with get_score(), chain_pairs, residues, write_outputs(), to_csv().

New capabilities:

• --csv export from CLI command
• Transparent gzipped-JSON support
• ~10x speedup on large complexes (AF2 example: 14.3s → 1.3s)

Bug fixes:

• Wrong "max" scores for non-alphabetical chain order: the original interleaved max computation with asym computation, so when chains appeared out of alphabetical order in the file, maxima were taken against not-yet-computed (zero) values. Now all asymmetric values are computed first, then maxima. This was verified with a synthetic test.
• Boltz crash on missing pair_chains_iptm: the fallback was dead code immediately followed by a direct dict access (KeyError). This now falls back cleanly to ipTM = 0.
• Companion-file path derivation (Boltz pLDDT/confidence, AF3 summary) used str.replace on the full path, breaking for directories containing "pae"/"confidences". Now basename-only.
• Clear errors instead of NameError/silent misbehavior for missing pae keys and PAE-matrix/structure size mismatches; .pkl PAE files now reachable.

Tests

• All three output files (.txt, _byres.txt, .pml) are byte-identical to the original script on both bundled examples, enforced by a 15-test pytest suite with fixtures generated from the original script.
• Code review feedback addressed (documented the intentionally preserved calc_d0_array L=27 quirk).
• Clean CodeQL scan.