On the Generalization of Equivariance and Convolution in Neural Networks to the Action of Compact Groups