Reducing self-supervised learning complexity improves weakly-supervised classification performance in computational pathology