Scalable spectral representations for multi-agent reinforcement learning in network MDPs