DFM-VLA: Iterative Action Refinement for Robot Manipulation via Discrete Flow Matching