Learning Diagonal Gaussian Mixture Models and Incomplete Tensor Decompositions