Add opt-in concurrent evaluation (overlap eval with next solve)

[robdmac]

Currently, evaluation runs serially between checkpoints - a checkpoint's tests must finish before the agent starts the next one even though eval results never feed back into the (blind) agent. This adds an opt-in --concurrent-evaluation flag that instead runs each checkpoint's eval concurrently with the next checkpoint's solve, so eval no longer blocks progress. On subprocess/compile-heavy problems (such as test_translator), evaluation is ~57% of per-run wall-clock so this approximately halves the run-time.

A single rolling background thread evaluates checkpoint N while N+1 solves; the previous eval is joined before the next starts, so solve stays <= 1 checkpoint ahead of eval and at most one solve + one eval run at a time (in-flight containers bounded to 2). Results land per-checkpoint as each eval completes, not batched at the end. Off by default.

Since eval never feeds the agent, scores are unchanged for the ANY_CASE pass policy, verified by re-evaluating recorded snapshots: pass counts and test_collection_hash bit-for-bit identical at checkpoints.

Trade-off: can't early-stop on test failures (a checkpoint's eval finishes during the next solve); agent errors and rate limits will stop the run.

[robdmac]

This can be hardened. Converting to draft.