FRIEDA: Benchmarking Multi-Step Cartographic Reasoning in Vision-Language Models