Not All Instances Are Equally Valuable: Towards Influence-Weighted Dataset Distillation