refactor(bench): reduce S3776 cognitive complexity across 7 benchmark files

Extract nested functions and complex inner loops to module-level helpers in
evaluator.py, multi_swebench.py, aggregate_sweep.py, aider_baseline.py,
bm25_baseline.py, diffctx_eval_fn.py, and calibrate.py to satisfy SonarCloud
S3776 threshold (≤15) without changing behaviour.

Author: nikolay-e