Learning Gaussian Graphical Models with Latent Confounders