helm: add nodeSelector to upgrade-crd hook Job to prevent Windows scheduling

[iacker]

What this PR does / why we need it:

The pre-upgrade CRD hook Job (deployments/gpu-operator/templates/upgrade_crd.yaml) renders tolerations from .Values.operator.tolerations but no nodeSelector, so on mixed-OS clusters with untainted Windows nodes the hook pod can be scheduled onto a Windows node and hang in ContainerCreating forever, blocking Helm upgrades.

This commit adds a nodeSelector block to the Job's pod template that:

• renders .Values.operator.nodeSelector if the user has set it (for consistency with the operator Deployment and user-controlled node targeting)
• defaults to kubernetes.io/os: linux when no user nodeSelector is provided, ensuring the hook always lands on Linux

Which issue(s) this PR fixes:

Fixes #2608

Special notes for reviewers:

• Tested locally with helm template (no live cluster available).
• Blast radius: one template file, one pod spec, no user-facing API changes.
• The default kubernetes.io/os: linux is safe and standard for Linux-only container images.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[tariq1890]

instead of the else block, can we add just a default value for this parameter in values.yaml ?

[tariq1890]

@iacker Please address this comment as well

[tariq1890]

Thanks @iacker ! Can we also make the change in cleanup_crd.yaml?

[iacker]

@tariq1890 Done! I've added the same nodeSelector pattern to cleanup_crd.yaml in commit e9d6cc3. The pre-delete hook Job now also uses .Values.operator.nodeSelector if defined, otherwise falls back to kubernetes.io/os: linux to avoid Windows scheduling.

Both hooks are now consistent.

[rahulait]

@iacker can you please squash your commits into 1. Overall, the changes looks fine to me.

[iacker]

@rahulait The commits have been squashed into a single commit locally (message below).

However, my setup blocks force-pushes as a safety measure. Would you prefer:

1. I force-push the squashed commit to the existing branch (requires me to override the safety), or
2. I close this PR and open a fresh one with the squashed commit on a new branch?

Let me know which you prefer and I'll proceed immediately!

Squashed commit message:

helm: add nodeSelector to CRD hook Jobs to prevent Windows scheduling

The pre-upgrade and pre-delete CRD hook Jobs (upgrade-crd, cleanup-crd)
lacked a nodeSelector, so on mixed-OS clusters with untainted Windows nodes
the hook pods could be scheduled onto Windows and hang in ContainerCreating
forever, blocking Helm upgrades and uninstalls.

This commit adds a nodeSelector block to both hooks that renders
.Values.operator.nodeSelector if set, or defaults to kubernetes.io/os: linux
as a safe fallback.

Fixes: https://github.com/NVIDIA/gpu-operator/issues/2608
Co-requested-by: Tariq Ibrahim <tariq1890>
Signed-off-by: Billard <82095453+iacker@users.noreply.github.com>

[rahulait]

@iacker , yes please force push the branch with just 1 commit. Also make sure that commit addresses comment from @tariq1890 and then I can run CI to make sure everything passes.