Exploiting Anonymity in Approximate Linear Programming: Scaling to Large Multiagent MDPs