Minimize the energy distance and use a better gradient penalty.

[fidlej]

I updated the code to minimize the energy-distance-based generator loss
when training the generator.
Previously, the code was training the generator to minimize the surrogate loss.

And I also used a better input to the gradient penalty,
from the updated paper:
https://openreview.net/forum?id=S1m6h21Cb

Thanks for the initial code.