Match Paper implementation closer

[jacobbieker]

Detailed Description

The paper's diagram of the model is this:

The implementation in this repo has three known differences.

1. There is only a single G block between each ConvGRU
2. Zsp input channels has to match the input channels from the conditioning stack, instead of the conditioning stack having half the channels of the latent space
3. The number of channels from the 3x3 conditioning stack outputs does not exactly match the paper, and changes the number of channels to match the diagram.

Context

These differences could be important in training the model. While the current implementation is quite close and almost entirely follows the paper, these differences could be quite important.