Easier install

[illeatmyhat]

Part of #2
This PR primarily focuses on providing a docker container that generally works, and makes the code as platform independent as possible with ROCm and CPU fallback.
Because requirements.txt was not installable as is, libraries were also updated, and deprecated functions replaced.

Particularly, those using VS Code and have Development Containers set up should be able to just open the folder and have a correctly running environment.

Caveats

1. PyTorch has no support for GPUs on Macs whatsoever.
2. CPU inference jobs are sometimes Killed in Docker because of additional memory overhead that tends to exceed the capacity of a Mac.
3. The output is distorted, maybe because of breaking changes in nn.Upsample and imresize()

[TsXor]

I don't know if you have used that maybe well-known DeepCreamPy, but your symptom looks relatively like "color bleeding".
In DeepCreamPy, marking the region you want to inpaint with brush tool, which almost always have anti-aliasing will cause such "color bleeding".
Look into this link for details: https://github.com/liaoxiong3x/DeepCreamPy/blob/master/TROUBLESHOOTING.md#strange-color-decensors

[TsXor]

According to the paper, you can output of every subroutine to locate the problem:

• ScreenVAE encoded map -> ScreenVAE Encoder network
• inpainted ScreenVAE map -> Semantic Inpainting network
• inpainted lines -> Semantic Inpainting network
• all the 3 above have no exception -> Appearance Synthesis network.

If the problem is really color bleeding, the exceptions should have happened in Semantic Inpainting network, and inpainted lines and ScreenVAE map should be abnormal.
What's more, you mentioned imresize(), and in another implementation by this author, he used ANTIALIAS param in imresize(), and as mentioned above, color bleeding is almost always caused by anti-aliasing.