create a docker container for reproducible model inference

[NiklasTR]

Hi Wengong,

I read the RefineGNN paper today and would love to run it within a docker container. I created a dockerfile that I hope fulfills the installation requirements. One thing I noticed was a difficulty around installing torch==1.8.2 - the long term support version seems to be not available via pip install. Have you seen this problem, too? I simply used torch==1.9.0 in this build - It seems like one might get more lucky using a conda version of pytorch-lts if 1.8.2 is a strict requirement.

How have you installing torch==1.8.2?