helm: add nodeSelector and tolerations to NIMCache templates

[abhay1999]

Summary

Fixes #1636

NIMCache resources were missing nodeSelector and tolerations fields across all NIM Operator mode Helm templates. Without these fields, NIMCache pods cannot be scheduled on GPU nodes that have taints or require node selection (e.g. cloud provider GPU node pools with nvidia.com/gpu: NoSchedule taints).

The NIMService sections of the same templates already exposed these fields correctly — this PR makes NIMCache consistent with NIMService.

Changes

Added nodeSelector and tolerations to the NIMCache spec in all 8 affected templates:

• helm/templates/llama-nemotron-embed-1b-v2.yaml
• helm/templates/llama-nemotron-rerank-1b-v2.yaml
• helm/templates/nemotron-graphic-elements-v1.yaml
• helm/templates/nemotron-ocr-v1.yaml
• helm/templates/nemotron-page-elements-v3.yaml
• helm/templates/nemotron-table-structure-v1.yaml
• helm/templates/nemotron-nano-12b-v2-vl.yaml (also adds missing nodeSelector to NIMService)
• helm/templates/nemotron-parse.yaml (also adds missing nodeSelector to NIMService)

Before / After

Before — NIMCache would ignore nodeSelector/tolerations from values.yaml, causing pods to be unschedulable on tainted GPU nodes:

kind: NIMCache
spec:
  source: ...
  storage: ...
  # nodeSelector and tolerations missing

After — consistent with NIMService:

kind: NIMCache
spec:
  source: ...
  storage: ...
  nodeSelector:
    {{ toYaml .Values.nimOperator.<model>.nodeSelector | nindent 4 }}
  tolerations:
    {{ toYaml .Values.nimOperator.<model>.tolerations | nindent 4 }}

Testing

Verified template structure matches the existing NIMService pattern used consistently across all templates in the chart.

[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.

[abhay1999]

Hi team! 👋 Following up on this PR — it's been 15 days since opening.

This adds nodeSelector and tolerations support to the NIMCache Helm templates, making it possible to schedule cache pods on specific nodes (GPU node pools, spot instances, etc.).

I noticed the copy-pr-bot flagged that this needs additional vetting before workflows can run on NVIDIA's runners. Happy to assist with anything needed from my side to unblock the validation. Thanks!