Fix multi-GPU runtime error with multi-scale netD

[cuihaoleo]

This PR solves #34 .

About the original issue: When torch.nn.DataParallel replicates a D_NLayersMulti object to multiple devices, although all submodules are replicated, self.model (which is an instance of ListModule) is not replicated because PyTorch doesn't know how to copy it correctly. As a result, all replicated D_NLayersMulti instances have self.model pointing to the same ListModule object whose module attribute points to original D_NLayersMulti object before replication. You can check id(self.model) in D_NLayersMulti.forward to verify it.

I removed the usage of ListModule class, and made D_NLayersMulti.forward directly call submodule.

[junyanz]

Thanks for the fix. Really helpful.