CI: Remove nccl installation after upstream fix (#18965)

https://gitlab.com/nvidia/container-images/cuda/-/issues/68

ci/docker/Dockerfile.build.centos7

@@ -91,32 +91,6 @@ RUN cd /usr/local/src && \
     cd /usr/local/src && \
     rm -rf ccache
 
-# NCCL is missing on CentOS7 images https://gitlab.com/nvidia/container-images/cuda/-/issues/68
-# Install manually if this is a GPU image; different Cuda versions require different NCCL versions
-# https://wiki.bash-hackers.org/syntax/pe#search_and_replace
-# We need to redeclare ARG due to
-# https://docs.docker.com/engine/reference/builder/#understand-how-arg-and-from-interact
-ARG BASE_IMAGE
-RUN export SHORT_CUDA_VERSION=${CUDA_VERSION%.*} && \
-    if [[ "$BASE_IMAGE" == *"nvidia/cuda"* ]]; then \
-        if [[ ${SHORT_CUDA_VERSION} == 9.2 ]]; then \
-            export NCCL_VERSION=2.4.8; \
-        elif [[ ${SHORT_CUDA_VERSION} == 10.* ]]; then \
-            export NCCL_VERSION=2.6.4; \
-        else \
-            echo "ERROR: Cuda ${SHORT_CUDA_VERSION} not yet supported in Dockerfile.build.centos7"; \
-            exit 1; \
-        fi && \
-        curl -fsSL https://developer.download.nvidia.com/compute/machine-learning/repos/rhel7/x86_64/nvidia-machine-learning-repo-rhel7-1.0.0-1.x86_64.rpm -O && \
-        rpm -i nvidia-machine-learning-repo-rhel7-1.0.0-1.x86_64.rpm && \
-        yum -y check-update || true && \
-        yum -y install \
-            libnccl-${NCCL_VERSION}-1+cuda${SHORT_CUDA_VERSION} \
-            libnccl-devel-${NCCL_VERSION}-1+cuda${SHORT_CUDA_VERSION} \
-            libnccl-static-${NCCL_VERSION}-1+cuda${SHORT_CUDA_VERSION} && \
-        yum clean all; \
-    fi
-
 # Fix the en_DK.UTF-8 locale to test locale invariance
 RUN localedef -i en_DK -f UTF-8 en_DK.UTF-8