Learning Additive Exponential Family Graphical Models via $\ell_{2,1}$-norm Regularized M-Estimation