[Feature]: Support affinity / nodeSelector for operand DaemonSets in ClusterPolicy spec.daemonsets

[yalattas]

Feature description

Add affinity and nodeSelector to the common DaemonSet config in ClusterPolicy
(spec.daemonsets), surfaced in the Helm chart as daemonsets.affinity /
daemonsets.nodeSelector (default empty). When set, apply them to all operand
DaemonSets (driver, container-toolkit, device-plugin, dcgm-exporter, gfd,
mig-manager, …), on top of the operator's existing NFD-based node selection.

Motivation

spec.daemonsets already exposes tolerations but has no affinity/nodeSelector.
That's an asymmetry: tolerations only permit scheduling onto tainted nodes — they
can't restrict where operands land. On a cloud cluster (EKS/GKE/AKS) with several
autoscaled node pools where only one has GPUs, I want to declaratively pin operands to
that pool via a label it already carries (e.g. karpenter.sh/nodepool=gpu).

The documented alternatives don't cover this well:

• nvidia.com/gpu.deploy.<operand>=disabled labels / node taints (how Deploy nvidia-device-plugin-daemonset to only certain nodes #547 was
  closed) are subtractive: you must label or taint every node you want to exclude.
  In autoscaled clusters (Karpenter, managed nodegroups) where nodes churn constantly,
  that doesn't scale and is race-prone.
• NFD gating keeps operands off non-GPU nodes, but can't target a specific GPU
  pool and doesn't apply when NFD is disabled.

A first-class affinity/nodeSelector is declarative, applies automatically to new
nodes, and is consistent with the operator's own NvidiaDriver CRD, which already
exposes spec.affinity and spec.nodeSelector.

Prior art

#547 requested essentially this and was closed pointing at the subtractive workarounds
above; a community MR to add nodeSelector to the common daemonset config was referenced
there but never landed.

Proposed API

# ClusterPolicy
spec:
  daemonsets:
    nodeSelector:            # proposed
      karpenter.sh/nodepool: gpu
    affinity: {}             # proposed

# Helm values.yaml
daemonsets:
  affinity: {}       # default: unchanged behavior
  nodeSelector: {}   # default: unchanged behavior

User-supplied values should be merged (AND) with the operator's NFD-based selection so
GPU detection still applies.

[amarkdotdev]

Opened #2645. Adds spec.daemonsets.affinity and nodeSelector with Helm wiring; user nodeSelector labels merge with the existing deploy labels.