Exchange-of-Thought: Enhancing Large Language Model Capabilities through Cross-Model Communication