What Makes a Good Unmasking Policy? A Controlled Study of Learned Unmasking in Diffusion Language Models