Probing the Limits of Stylistic Alignment in Vision-Language Models