Text this: Training a high-performance retinal foundation model with half-the-data and 400 times less compute