Learning from Complexity: Exploring Dynamic Sample Pruning of Spatio-Temporal Training