fix: restrict KEDA scaling to /v1/(chat/)?completions paths only

- Switch to HTTPScaledObject (matches a100 pattern)
- Probes (/health, /v1/models) now ignored for scaling
- Keeps fast polling (10s), quick cooldown (30s), 15min idle scale-to-0

Health/readiness/startup probes hit pod IP directly (bypass service), so never triggered scaling anyway — but paths ensure only inference traffic scales.

infrastructure/gpus/base/keda-gpu-scaling/keda-vllm.yaml

@@ -1,17 +1,22 @@
-apiVersion: keda.sh/v1alpha1
+apiVersion: http.keda.sh/v1alpha1
-kind: ScaledObject
+kind: HTTPScaledObject
 metadata:
   name: rtx6000-scaling
   namespace: customer1
 spec:
+  hosts:
+  - rtx6000-brain-service.customer1.svc.cluster.local
   scaleTargetRef:
     name: rtx6000-brain-vllm
     kind: Deployment
     apiVersion: apps/v1
-  hosts:
+    service: rtx6000-brain-service
-    - rtx6000-brain-service.customer1.svc.cluster.local
+    port: 8000
+  paths:
+  - path: "^/v1/(chat/)?completions"
   replicas:
     min: 0
     max: 1
-  targetPendingRequests: 1
+  pollingInterval: 10
-  scaledownPeriod: 900\n  pollingInterval: 10\n  cooldownPeriod: 30
+  cooldownPeriod: 30
+  scaledownPeriod: 900