Fair Diffusion: Instructing Text-to-Image Generation Models on Fairness