Evaluating LLMs on Generating Age-Appropriate Child-Like Conversations