Learning to Negotiate: Multi-Agent Deliberation for Collective Value Alignment in LLMs