Learning without relearning physical symmetry: an SO(3)-equivariant convolutional network (CNN) on the sphere