Judge tuning (4/5): Run multi-fidelity judge searches

[ErlisLushtaku]

Runs the multi-fidelity search. Each NePS trial writes a config for the target meta-eval task and runs it in a subprocess, so the GPU is freed between judges and every trial is a normal meta-eval run that shares its judgement cache.

• Objectives are agreement (NePS minimizes 1 - agreement) and expected judging cost in USD per 1,000 battles, computed from native token counts including cached judgements. NePS's own cost field stays unset because it measures search budget, not judge cost.
• Failed trials, such as a CUDA out-of-memory, are recorded with infinite objective values and the search continues (ignore_errors defaults to true), since one bad configuration should not stop a long run. Pricing and config errors still abort.
• NePS returns once the budget is spent, possibly while other workers are still evaluating. The primary process therefore waits for active trials, reports the agreement/cost Pareto front, picks the best agreement configuration per judge model at full fidelity and evaluates those on the test split. Helpers join with --tune_judge.search_only true and the same run_dir.
• Prices come from tune_judge.price_per_million_tokens, and a cost run stops before evaluating if a searched model has none. Automatic reference prices follow in Judge tuning (5/5): Add reference prices and usage docs #136.

Merge order: #138 → #135 → #140 → #133 → #136.