ESP: Exploiting Symmetry Prior for Multi-Agent Reinforcement Learning