What Do Chart Question Answering Benchmarks Measure? Task-Specific Visual Evidence Dependence in Vision-Language Models