Evaluation: Worker cap for LLM jobs

Is your feature request related to a problem?
Fast evaluations and high-priority LLM jobs share one Celery default queue and one global worker_concurrency, with no cap on how many fast-eval chunk tasks can run at once. A 500-item run fans out to 10 chunk tasks (chunk size 50), so 2-3 concurrent eval runs (20-30 tasks) can occupy every worker slot.

Task priority (LLM=9, eval=6) does not help here: RabbitMQ priority only decides which queued task runs next when a slot frees — it does not preempt a running task. So when all slots are busy with eval chunks, an incoming LLM job waits until a chunk finishes (soft time limit 300s), meaning a high-priority LLM call can wait ~4 minutes. Not acceptable for interactive latency.

Describe the solution you'd like
Cap the worker capacity fast evals can consume, so high-priority LLM jobs always have headroom