Emergence of Biased Consensus in Multi-Agent LLM Debates