Scalable Multi-Agent Reinforcement Learning for Networked Systems with Average Reward