Standardize the module structure. Make the list of requirements more complete.

[Poikilos]

• Make the repo structure use Python standards to make all scripts work (tested with Python 3.9).
• Make it work with new scikit-image: ed1c575
• Document issues regarding specific versions of:
  • CUDA compute capability
  • CUDA Toolkit
  • NVIDIA driver
  • pytorch
  • torchvision

I had to do all of this just to get the tests.sh command in the readme to run, so this PR is probably necessary for maintenance. Let me know if you have any comments or questions.

The specific older versions of pytorch and torchvision in my new requirements.txt are probably fine since:

• The project itself is rather old in terms of ML.
• The versions I set are stated to work with cards with CUDA compute capability 3.7, and with the state of GPU prices and electronic waste, people can't be expected to buy expensive newer cards when older high-end GPUs work fine. If there are problems with these versions let me know.

Other changes:

• Add the GitHub-recommended Python.gitignore and add additional lines specific to the project to make commits, branching, forking, testing, etc much cleaner (ignore Python-generated and project-generated data).

Alternative PR (not initiated):
If you like pgd (I'm not familiar with it) as shown in this other fork: hyunobae@b79afe9 then I have a separate branch for that you can pull instead, so let me know or go ahead and start a new PR with my pgd-and-new-sckitit-image branch instead: https://github.com/poikilos/TecoGAN-PyTorch/tree/pgd-and-new-scikit-image