LensVLM: Compressing long context as images, expanding only relevant pages