How to organize queue, rate limiting, and request handling to LLM when the model's available throughput is exceeded?