Dissecting Chain-of-Thought: A Study on Compositional In-Context Learning of MLPs