Multi-Persona Thinking for Bias Mitigation in Large Language Models