Create compatibility matrices manually

[bfirsh]

We have a script that scrapes the PyTorch and Tensorflow websites to generate what versions of Torch/Tensorflow/CUDA/etc are compatible with each other.

There are a few problems with this:

• It is extremely fragile and breaks often, but the entries for previous versions never change so there's no need to keep on editing past versions.
• There are a few entries that are incorrect on PyTorch/Tensorflow's website.
• Some are just missing.

Torch and Tensorflow versions don't get released very often. It would be less work to just edit the compatibility matrix manually than to fix our scraping script.

Perhaps we just manually add new Torch/Tensorflow versions. Or, perhaps semi-manually curate it where we have a script that adds the template of a new version and we manually update it to ensure it's correct.

See also #569

[tempusfrangit]

With the addition of #2400 we have materially improved compatgen. The CuDNN is not extracted from the Docker image config environment variables (NV_CUDNN_VERSION) instead of parsing the tag string.