Revealing Political Bias in LLMs through Structured Multi-Agent Debate