Learning a Gaussian Mixture for Sparsity Regularization in Inverse Problems