Implement utilities to train generative models for Scarlet

[EiffL]

This issue is to keep track of the development necessary to train generative models for use with Scarlet for deblending.

• [x] : Implement PixelCNN model in g2g
• [x] : Define training set properties from GalSim COSMOS dataset
• [ ] : Train model
• [ ] : Add export scripts to make a scarlet friendly tf hub module

Just to keep track of things, this issue is connected to the scarlet implementation of deblending with ML learned priors here: https://github.com/pmelchior/scarlet/issues/128

I'm opening branch u/EiffL/pixelcnn to work on it.

[EiffL]

I've added code to generate cosmos postage stamps at HSC resolution, and implemented the pixel cnn model. Note that there might still be some work needed on the effective PSF used for cosmos, generated here https://github.com/ml4astro/galaxy2galaxy/blob/u/EiffL/pixelcnn/notebooks/BuildingCOSMOS_PSF.ipynb There are traces of aliasing in some images...