Support different cuda versions in Python Predictor base image

[deliahu]

Motivation

As shown in TensorFlow's compatibility matrix, different versions of TensorFlow rely on different versions of CUDA. The same is true for Pytorch.

Solution

Build multiple versions of python-gpu-slim:

cortexlabs/python-predictor-gpu-slim:<cortex_version>-cuda10.0
cortexlabs/python-predictor-gpu-slim:<cortex_version>-cuda10.1
cortexlabs/python-predictor-gpu-slim:<cortex_version>-cuda10.2
cortexlabs/python-predictor-gpu-slim:<cortex_version>-cuda11.0

And for the full version of the image python-gpu:

cortexlabs/python-predictor-gpu:<cortex_version>

Here is an example of using a Docker build arg:

ARG CUDA_VERSION=10.1
FROM nvidia/cuda:$CUDA_VERSION-cudnn7-devel-ubuntu18.04