Proposal: Add generic file for NNet

[mikhail]

Just want to start a conversation here -- this is an example of how one generic NNet file could exist (example in pytorch) ( #194 )

The only thing the individual games would have to specify is number of layers.

This is of course an oversimplification because all layers would be identical (same stride/padding) but this can be configurable as well.

[mikhail]

The main goal here is to provide a developer of a new game with an easy way to configure the learning layers. Some sort of abstraction layer where they could simply provide a configuration of something like this:

nnet_settings = {
    'layers': 10,
    'nnets': ['conv2d', 'batchnorm', 'linear']  # didn't do this
}

That's it. The abstraction layer would do the rest.

[mikhail]

@suragnair This is pretty close to where I wanted to have it. Take a look?

[mikhail]

@suragnair this is working now. Would love your thoughts on this generalization