Towards Efficient LLM Grounding for Embodied Multi-Agent Collaboration