Induction in Both Directions: A Mechanistic Analysis of In-Context Learning in Masked Diffusion Language Models