Vault: Augmenting the vision-and-language transformer with the propagation of deep language representations