Multi-role Consensus through LLMs Discussions for Vulnerability Detection