Disentangling the Gauss-Newton Method and Approximate Inference for Neural Networks