Text this: Understanding pre-training data effects in retinal foundation models using two large fundus cohorts.