Boosting the Cross-Architecture Generalization of Dataset Distillation through an Empirical Study