Understanding Alignment in Multimodal LLMs: A Comprehensive Study