Refresh workload, RoCE, and registry examples

[dholt]

Summary

• Refresh active workload and container examples to current NGC images: RAPIDS/Dask notebooks to nvcr.io/nvidia/rapidsai/notebooks:26.04-cuda12-py3.13, PyTorch Kubernetes smoke job and Slurm validation base container to current NGC tags, and the rootless Docker smoke to nvcr.io/nvidia/cuda:13.0.2-base-ubuntu24.04.
• Update the RAPIDS example Dockerfile and prepare.sh for the current notebooks image layout (notebooks under /opt/rapids/notebooks).
• Refresh RoCE backend defaults and docs to MLNX_OFED_LINUX-24.10-4.1.4.0-ubuntu24.04-x86_64 and current k8snetworkplumbingwg Multus/SR-IOV manifest URLs; document that new Kubernetes RDMA/RoCE deployments should prefer the NVIDIA Network Operator, with the DeepOps role retained as a legacy direct-host path.
• Replace the stale public NCCL test image example with site-owned image placeholders and update NGC Ready defaults/docs for PyTorch and TensorFlow.
• Update registry-cache defaults to include registry.k8s.io and correct the documented registry proxy image to rpardini/docker-registry-proxy:0.6.5.

Validation

• git diff --check origin/master..HEAD
• YAML parse for all changed YAML files
• Ansible syntax-check suite plus focused checks for playbooks/k8s-cluster/roce.yaml and playbooks/slurm-cluster/slurm-validation.yml
• Docker manifest inspection for the referenced NGC/CUDA/RAPIDS image tags and HTTP verification of the OFED download URL
• OS compatibility audit against this branch: high=0, medium=0, low=45, info=30
• Public sanitizer check on the PR body

Notes

• This is an examples/docs/defaults refresh; it does not change deployment playbook behavior.
• The RoCE backend role is kept as a legacy direct-host path; Network Operator remains the recommended route for new Kubernetes RDMA deployments.