LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior