Use D-Adaptation for all DL models

[jakobnissen]

Svetalana uses D-Adaptation for the taxonomy predictor. Let's see if it's also good for the VAE and the other models.

• Use D-Adaptation and verify performance is the same or better
• Make setting the lrate a noop (or completely delete the flag, if before a major release)
• In the presence of D-Adaptation, recheck optimal starting batch size and batch steps (Experiments in optimising batch steps #180)