LensVLM: Selective Context Expansion for Compressed Visual Representation OfText