MAGIC-VQA: Multimodal And Grounded Inference with Commonsense Knowledge for Visual Question Answering