Cohorts: run_cohort — N subjects per run

[MuteJester]

Summary

Adds Experiment.run_cohort(genotypes, ...) — run N subjects, each with its own diploid Genotype, in one call. It is a Python loop around the existing single-subject genotype path (with_genotype → compile → run_records): each subject is compiled and run independently, records are tagged with subject_id and given a namespaced sequence_id, and the per-subject SimulationResult (with its own refdata) is collected into a new CohortResult. Follow-on #3 of the genotype roadmap.

Python-only — zero engine_rs/ changes. No new engine behavior; orchestration over already-correct per-subject behavior.

What's included

• Experiment.run_cohort(genotypes, *, n_per_subject=1, seed=0, counts=None, strict=False, expose_provenance=False, validate_records=False, allow_curatable_refdata=None) — up-front validation, deterministic per-subject sub-seeds, clone-per-subject (copy.copy, never mutates the base experiment), per-subject with_metadata stamping.
• src/GenAIRR/cohort.py — CohortSubjectResult (frozen: subject_id, genotype, result, refdata, seed, count) and CohortResult: derived .subject_ids/.genotypes/.results, .result_for/.refdata_for, independent .records, combined .to_dataframe/.to_csv/.to_fasta (union columns; unique >{sequence_id} FASTA headers), __len__. Per-subject refdata stored explicitly (which SimulationResult does not preserve).
• Counts: n_per_subject for all, or a parallel counts list for varied repertoire sizes (incl. 0 — subject present with zero records). Mappings/str/bytes rejected.
• Subject IDs: taken from each genotype, or auto-assigned subject_0..N-1; mixed/duplicate (incl. 1 vs "1") raise. Resolved on snapshots only — caller genotypes never mutated.
• Docs: new "Cohorts" guide section; removed from Limitations.

Behavior

• Mutually exclusive with with_genotype, restrict_alleles, recombine(*_allele_weights=...); not combined with receptor_revision/clonal forks (parity with the single-subject genotype path). Reserved metadata keys (subject_id/sequence_id/haplotype) raise.
• Per-subject refdata preserved → validate_records and novel-allele truth calls correct even when subjects differ.
• Deterministic: per-subject sub-seeds derived from seed.

Test plan

• New tests/test_genotype_cohorts.py (21 tests): CohortResult accessors + independence, union/FASTA export, subject-id + counts resolution, happy path, determinism, no-base-mutation (incl. mid-loop failure), counts override/zero, mutual exclusions, cartridge-hash, mixed novel/plain subjects + per-subject refdata, per-subject validate, with_metadata parity + collision guard, empty-export schema, end-to-end.
• Full Python suite green (2446 passed).
• No engine_rs/ changes.
• Two review rounds (spec-compliance + critic, both passes); findings fixed with regression tests.