Not All Samples Should Be Utilized Equally: Towards Understanding and Improving Dataset Distillation