Location-Aware Visual Question Generation with Lightweight Models