In-memory judge concurrency queue is a no-op on Workers (misleading)

**Severity:** Low (cosmetic) `src/lib/judge.ts:19-44` keeps a module-level `activeJobs`/`jobQueue` semaphore. Workers isolates don't share state across requests and get recycled, so this never enforces a real cluster-wide concurrency cap and `judgeQueueStats()` reports misleading numbers. Real concurrency limiting must live in the Rust/Akiro judge itself. The code comment already acknowledges this. **Fix** Either remove the dead limiter (and `judgeQueueStats`) to avoid confusion, or clearly document it as dev-only and never surface its stats in prod. **Acceptance:** no misleading "active/queued" numbers can reach production UI.