SEAP: Training-free Sparse Expert Activation Pruning Unlock the Brainpower of Large Language Models