QG-CoC: Question-Guided Chain-of-Captions for Large Multimodal Models