POEM: Interactive Prompt Optimization for Enhancing Multimodal Reasoning of Large Language Models