A Foundation LAnguage-Image model of the Retina (FLAIR): Encoding expert knowledge in text supervision