🐛 Describe the bug

Please see:
#149540

After landing of: #149351

CUDA aarch64 builds where broken, since nccl was defined in .ci/docker/common/install_cuda_aarch64.sh as well as in the matrix.

Consider removing the installation of nccl in the docker.
Change to have only 1 source of truth for both builds.

Versions

2.8.0

cc @seemethere @malfet @osalpekar @pytorch/pytorch-dev-infra