Fix dcgan_mnist model

[ghost]

Fixes:

• Replace Flux.logitbinarycrossentropy with Flux.Losses.logitbinarycrossentropy
• Take the latent_dim as input to the generator constructor

No yet fixed:

• Crashes, likely because of this issue : ConvTranspose on GPU fails with certain activation functions Flux.jl#1350

see #266

[ghost]

I removed the tanh activation of the last ConvTranspose layer and added it as an seperate layer, this works around the issue that triggered the crash. Scripts runs without problems now, output after 20 epochs seems ok.

[CarloLucibello]

I think another manifest got caught in the update, it doesn't really matter although it will conflict with the other PR.
I'll merge whenever you say this is ready in any case

[ghost]

Ah yes, I missed the extra 'c' in the filename. The cdcgan models runs fine with the updated manifest.

The biggest problems with this model are solved I think, so LGTM