Dataset Distillation with Neural Characteristic Function: A Minmax Perspective