Add discriminative regularization to mlp output

[dribnet]

Mostly resolves #5.

@vdumoulin - curious if this looks roughly right. I've confirmed that training starts, but haven't fully trained a model with this change yet. Can do so if this doesn't have any obvious flaws and that would be useful.

Note: also includes a commit to change model loading so that this can be run with more recent vintage blocks.

[dribnet]

Update: this runs and trains, but the loss explodes after about 6 epochs - and I've confirmed it's the new discriminative mlp term causing this. So something here is not quite right.