LLM+P: Empowering Large Language Models with Optimal Planning Proficiency