Implicit Bias of Gradient Descent on Linear Convolutional Networks