MMRefine: Unveiling the Obstacles to Robust Refinement in Multimodal Large Language Models