On the Representational Capacity of Neural Language Models with Chain-of-Thought Reasoning