A Universal Load Balancing Principle and Its Application to Large Language Model Serving