ImageChain: Advancing Sequential Image-to-Text Reasoning in Multimodal Large Language Models