ci-wheel: ensure libnccl is always installed

[jameslamb]

While working through CUDA 13.1 rollout in RAPIDS (rapidsai/build-planning#236), we noticed that wheel builds requiring libnccl were failing (example: rapidsai/raft#2912 (comment))

The root cause appears to be:

• those builds depended on finding a system-installed libnccl
• the available CUDA 13.1 nvidia/cuda images (base images used here) didn't come with it preinstalled

This PR proposes installing libnccl-dev(el) at build time of rapidsai/ci-wheel if it isn't found from the base image.

Notes for Reviewers

How I tested this

Manually pushed similar changes to the build scripts in the RAFT PR where these failures were observed, and saw builds succeed: rapidsai/raft#2912

Confirmed that I saw the expected log messages and image contents from builds here. For example, in the build-images / ci-wheel / build (13.1.0, 3.13, rockylinux8, amd64) build (link)

#9 37.11 libnccl-devel not found, manually installing it
...
#9 38.23 Installing:
...
#9 38.23  libnccl-devel             x86_64  2.28.9-1+cuda13.0                       cuda        63 k

and

docker run \
  --rm \
  --pull always \
  docker.io/rapidsai/staging:ci-wheel-342-26.02-cuda13.1.0-rockylinux8-py3.13-amd64@sha256:ba0dadd01dc5145b280966b54b72e3ef86cd114b08b9c12feb867b0cb43c2aa1 \
  find / -name 'libnccl.so*'

# /usr/lib64/libnccl.so.2.28.9
# /usr/lib64/libnccl.so
# /usr/lib64/libnccl.so.2

[bdice]

Maybe we can use the same array logic in other Dockerfiles? Let's do that as a followup.

[jameslamb]

sure, happy to do a small followup soon with this and some other minor cleanup