Do Large Language Models Have Compositional Ability? An Investigation into Limitations and Scalability