Edge-free but Structure-aware: Prototype-Guided Knowledge Distillation from GNNs to MLPs