Reasoning Paths with Reference Objects Elicit Quantitative Spatial Reasoning in Large Vision-Language Models