Making LLMs Cheaper and Better via Performance-Efficiency Optimized Routing