Add quantization aware training to new version of the code

[zetyquickly]

Initial version of the repo has possibility to quantize neural network. After vast refactoring done by @rakhimovv to address the problem that results from the paper cannot be reproduced, we've lost quantization functionality.

It seems like community is interested (link1 link2) in QAT present in previous version.

With issue I'd like to propose you to contribute and re-implement this functionality