Learning Heteroscedastic Models by Convex Programming under Group Sparsity