Visually Grounded Active Perception VLM for Ultra-High-Resolution Remote Sensing Understanding