The check fails in recent CI, but I think it should be correct
anyway. Also, I dont think that error bounds below single precision
accuracy are good for numdiffs.