Improving Zero-Shot Models with Label Distribution Priors