MASSV: Multimodal Adaptation and Self-Data Distillation for Speculative Decoding of Vision-Language Models