Towards Efficient Deep Spiking Neural Networks Construction with Spiking Activity based Pruning