Multi-Task Learning for Visually Grounded Reasoning in Gastrointestinal VQA