Few-Shot VQA with Frozen LLMs: A Tale of Two Approaches