Towards Detecting LLMs Hallucination via Markov Chain-based Multi-agent Debate Framework