Scalable and Accurate Graph Reasoning with LLM-based Multi-Agents