Synthesizing Post-Training Data for LLMs through Multi-Agent Simulation