Efficient Bound of Lipschitz Constant for Convolutional Layers by Gram Iteration