FOCUS: Towards Universal Foreground Segmentation