Leveraging Partial Symmetry for Multi-Agent Reinforcement Learning