Did Translation Models Get More Robust Without Anyone Even Noticing?