Practical attention implementations don't work over arbitrary length sequences. The universal approximation theorem holds IMO. Information will mix as you go through fully connected MLP layers. Attention is apparently a prior structure that is needed to really reduce training costs.