Revisiting Random Channel Pruning for Neural Network Compression