The OS updates for Distributed machine computation on GPUs

[starlyxxx]

We now use Ubuntu 16.04 and Python 3.6 for GPU-based computation. Ideally, its environment can be further updated to Ubuntu 20.04 and Python 3.8 just like the environment of CPU-based computation.

The problem is that the GPU-based environment is more complicated. The corresponding CUDA and other libraries (like openmpi, libnccl) also need to be updated with the OS update. See details in Dockerfile.

Please make sure every library is compatible and working before the OS update of GPU-based computation.