HARP (Human-Assisted Regrouping with Permutation Invariant Critic): A Multi-Agent Reinforcement Learning Framework for Improving Dynamic Grouping and Performance with Minimal Human Intervention