Knowledge Integration for Physics-informed Symbolic Regression Using Pre-trained Large Language Models