Lightweight Vision-Language Models for Multimodal Indonesian Document Search