Can MLLMs Perform Text-to-Image In-Context Learning?