Update driver version of EC2 image

[dsuess]

Description

The current docker-images use CUDA 10.1, since that seems to be the least common denominator between all the packages used. However, when we tried to run pytorch with CUDA 10.2 (necessary because of pytorch extensions compiled with 10.2), we're getting a runtime error

RuntimeError: cuda runtime error (35) : CUDA driver version is insufficient for CUDA runtime version ...

Motivation

Pytorch 1.5 and 1.6 come by default compiled against CUDA 10.2 so it would be nice if the underlying EC2 image's driver would be compatible with CUDA 10.2.

[deliahu]

@dsuess do you mind sharing how you approached installing CUDA 10.2? The way I was thinking it could work would be to clone the cortex repo and modify the base cuda image used in the first line of images/python-predictor-gpu/Dockerfile. You also might as well re-purpose this section to install whichever packages you need (remove the test "${SLIM}" = "true" part as well), or remove the section entirely if you plan on using a requirements.txt. You can then run docker build -f images/python-predictor-gpu/Dockerfile from the root of the repo to build the image.

I also spoke with @RobertLucian, and we came up with plan where we could build and publish a base image of python-predictor-gpu-slim for each CUDA version. I updated #923 accordingly.