On the convex formulations of robust Markov decision processes