Add support for PyTorch

[mlajtos]

Generating executable code is a top priority. PyTorch is currently the top DL framework (because of the define-by-run nature), so naturally it would be very suiting to combine the two.

Moniel should provide easy way to define weird architectures and PyTorch is cabable of representing them. However, user must be able to interfere with the generated code, because high-level description often does not capture all the neccesities of the idea.

[mlajtos]

Since code is the best representation for computational graphs, output should be generated Python code using PyTorch infrastructure. Again, output WILL NOT be any JSON/YAML/PB stuff, but actual runnable code.

Since user will be able to change generated code (or maybe inject it from Moniel?), result will be the best of the both worlds.

[mlajtos]

There is a proof-of-concept code that constructs code that resembles PyTorch, but there are a few questions that need to be discussed:

• Generate torch.nn.Modules or just dump everything in one block of code?
• How to handle Tensors in a functional way? Something like tf.get_variable..?
• ...more to come