Fix ACGAN training

[yang-song]

The ACGAN training procedure is not optimal

[ElApseR]

I think this is correct and makes training stable.

By the article,

$$L_s=E[logP(S=real|X_{real})]+E[logP(S=fake|X_{fake})]$$ $$L_c=E[logP(C=c|X_{real})]+E[logP(C=c|X_{fake})]$$ $$D = argmax(L_s+L_c)$$ $$G = argmax(L_c-L_s)$$

which means, you don't need to train Q separately. Also, Q shouldn't optimize variables of G.