bench: parallelize + resume the Curve-B re-eval (HP-search sensitivity)

[davorrunje]

Context

PR #117's Curve B (test metric of the running incumbent) is reconstructed by re-evaluating each best-so-far changepoint config on the test set via final_eval. Today this re-eval is serial in a single process: it is launch-bound on the tiny datasets (one CPU core pegged, GPUs ~idle), and it writes its output JSON only at the very end (no checkpoint, so a killed run restarts from scratch). Because of the cost, Curve B is omitted for compas and loan (# re-eval = 0 in the docs saturation table).

Goal

Make the Curve-B re-eval (a) parallel across the device pool and (b) resumable, so it is ~10–20× faster and restartable — then regenerate compas and loan Curve B.

Where

• benchmarks/_common/sensitivity_report.py:incumbent_test_curve — the serial per-incumbent final_eval loop (a per-incumbent progress log already exists via progress_label).
• benchmarks/sensitivity.py — extract_dataset, _cmd_run.
• benchmarks/_common/gpu_pool.py:fan_out — the shared work-stealing subprocess pool to reuse.

Proposed approach

Emit the independent incumbent re-evals (28 per dataset here; or per-(dataset,flavor)) as jobs and fan them across GPU slots via fan_out (oversubscribe — nets are ~0.7 GB). Memoize each incumbent's trial_idx → test IQM to a small cache file so a restart skips completed ones. Then run compas/loan and commit their curve JSONs + the updated figure.

Acceptance criteria

• Curve-B re-eval runs concurrently across ≥2 GPU slots (GPU utilization no longer ~single-stream).
• A killed+restarted re-eval resumes (completed incumbents not recomputed).
• benchmarks/results/alternate-base/curves/{compas,loan}.json contain Curve B; figure regenerated; docs "omitted for compas/loan" note removed.

References

• PR feat(bench): HP-search sensitivity curves (spec + plan + implementation) #117; docs docs/benchmarks/hp-search-sensitivity.md ("Follow-ups").