Design Your Own Universe: A Physics-Informed Agnostic Method for Enhancing Graph Neural Networks