Text this: Benchmarking foundation models as feature extractors for weakly supervised computational pathology.