fix: HighNodeUtilization plugin incorrectly evicts all pods when only GPU resources are configured

[ditingdapeng]

What version of descheduler are you using?

descheduler version: main branch (latest development version)

Does this issue reproduce with the latest release?
Yes, this issue can be reproduced with the latest release.

Which descheduler CLI options are you using?
Using descheduler as a Kubernetes deployment with a ConfigMap-based policy configuration.

Please provide a copy of your descheduler policy config file

apiVersion: descheduler/v1alpha2
kind: DeschedulerConfiguration
kind: DeschedulerConfiguration
deschedulerPolicy:
  strategies:
    HighNodeUtilization:
      enabled: true
      params:
        thresholds:
          "cloudml.gpu/v100-32g": 40
          "cloudml.gpu/h20-96g": 40
        numberOfNodes: 0
  nodeSelector:
    node.miks.io/type: "gpu"

What k8s version are you using (kubectl version)?

Client version: v1.24.0

Server version: v1.21

What did you do?

1. Deployed descheduler with the above configuration which only includes GPU resource thresholds
2. Created pods with and without GPU resource requests on nodes
3. Observed the behavior when node GPU utilization falls below the threshold (40%)

Steps to reproduce:

# 1. Deploy a pod with GPU request
apiVersion: v1
kind: Pod
metadata:
  name: gpu-pod
spec:
  containers:
  - name: gpu-container
    image: nvidia/cuda
    resources:
      requests:
        "cloudml.gpu/v100-32g": "1"
      limits:
        "cloudml.gpu/v100-32g": "1"

# 2. Deploy a pod without GPU request
apiVersion: v1
kind: Pod
metadata:
  name: non-gpu-pod
spec:
  containers:
  - name: nginx
    image: nginx

What did you expect to see?
When a node's GPU utilization falls below the threshold:

• Only pods that have requested the specified GPU resources (cloudml.gpu/v100-32g or cloudml.gpu/h20-96g) should be considered for eviction
• Pods that don't use these GPU resources should not be affected

What did you see instead?
When a node's GPU utilization falls below the threshold:

• The HighNodeUtilization plugin evicts all evictable pods from the node, regardless of whether they use the specified GPU resources or not
• This includes pods that don't request any GPU resources, which should not be affected by the GPU utilization threshold

This behavior is problematic because:

1. It unnecessarily disrupts pods that don't contribute to the GPU utilization
2. It may cause service disruption for non-GPU workloads
3. It doesn't align with the principle of only balancing the specifically configured resources

The root cause appears to be in the pod filtering logic of the HighNodeUtilization plugin, where it doesn't properly check whether pods are using the configured GPU resources before selecting them for eviction.

[googs1025]

same as: #1634