Exploring Data Efficiency in Zero-Shot Learning with Diffusion Models