How Well Can Transformers Emulate In-context Newton's Method?