Reasoning in Conversation: Solving Subjective Tasks through Dialogue Simulation for Large Language Models