SCITUNE: Aligning Large Language Models with Scientific Multimodal Instructions