NORM: Knowledge Distillation via N-to-One Representation Matching