Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions