Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models