Add ATNs for ImageNet dataset

This issue is a continuation of #19. The code for MNIST will be merged in #85. The original paper also mentions architectures for ImageNet models. It would be nice to have these implementations in the repository.

In this [comment](https://github.com/Adversarial-Deep-Learning/code-soup/pull/85#issuecomment-923621353), I have added dummy code for all architectures needed. However, the implementation should contain exactly the same number of channels/kernel sizes as the original model.

This will need effort in checking the what input/output paddings and strides will work given the architecture.