[EACL 2026 Main] SCoPE VLM: Selective Context Processing for Efficient Document Navigation in Vision-Language Models