Text this: Revisiting model scaling with a U-net benchmark for 3D medical image segmentation