PreScope: Unleashing the Power of Prefetching for Resource-Constrained MoE Inference