Latent Agents: A Post-Training Procedure for Internalized Multi-Agent Debate