On the Limitation of Diffusion Models for Synthesizing Training Datasets