Parts of Speech–Grounded Subspaces in Vision-Language Models