One Model, Many Graphs: Learning over Attributed Graphs across Heterogeneous Modalities with Vision-Language Models