GFlowVLM: Enhancing Multi-step Reasoning in Vision-Language Models with Generative Flow Networks