Use the pytorch/builder script to install CUDA packages in nightly docker.

[xuzhao9]

Currently, we are building nightly dockers by copying the code from pytorch/builder.

In order to make sure the CUDA and cudnn versions are consistent with the upstream, we should use pytorch/builder scripts to install CUDA packages in the docker build.

The source of the pytorch/builder code to install cuda: https://github.com/pytorch/builder/blob/main/common/install_cuda.sh

The source of docker script installing cuda: https://github.com/pytorch/benchmark/blob/main/docker/gcp-a100-runner-dind.dockerfile#L34

In the dockerfile, we should checkout the pytorch/builder git repo and use install_cuda.sh to install the CUDA libraries.

We also want to keep using the latest CUDA version in install_cuda.sh by default, but that could be the next step.

[xuzhao9]

Fixed by #2138