Universal Model Routing for Efficient LLM Inference