A Benchmark for Counterfactual Evaluation of Reasoning Faithfulness in Medical Vision-Language Models