feat: add HPA autoscaling for InferenceService

[Defilan]

Summary

Adds native Kubernetes HPA (autoscaling/v2) support to InferenceService. This was the single biggest functional gap vs every competitor (KubeAI, KAITO, KServe, llm-d, AIBrix). When an autoscaling block is configured, the controller creates and manages an HPA targeting the inference Deployment.

Usage:

spec:
  modelRef: my-model
  replicas: 1
  autoscaling:
    minReplicas: 1
    maxReplicas: 5
    metrics:
      - type: Pods
        name: "llamacpp:requests_processing"
        targetAverageValue: "2"

CLI:

llmkube deploy llama-3.1-8b --gpu --max-replicas 5

How it works:

• Controller creates an HPA v2 resource owned by the InferenceService (garbage collected on delete)
• Default metric: llamacpp:requests_processing with target average of 2 per pod
• When HPA is active, controller sets Deployment replicas to nil (lets HPA manage)
• Metal accelerator workloads skip HPA (no Deployment to scale)
• Requires Prometheus Adapter for custom metrics (documented)

Design decisions:

• Native HPA v2, not KEDA (CNCF-aligned, no external dependencies)
• Structured autoscaling block (extensible for future behavior policies, scale-to-zero)
• MinReplicas minimum is 1 (no scale-to-zero in v1)

Changes

• 967 lines added across 8 files
• 11 new tests (unit + envtest), controller coverage 78.9% -> 80.3%
• CRD: AutoscalingSpec, MetricSpec types
• Controller: reconcileHPA(), constructHPA(), replica conflict avoidance
• CLI: --min-replicas, --max-replicas, --autoscale-metric flags
• Helm: autoscaling RBAC in ClusterRole

Test plan

• make test passes (186 tests, 11 new)
• Deploy with --max-replicas 3, verify HPA created
• Deploy without autoscaling, verify no HPA
• Remove autoscaling from spec (unit test), verify HPA deleted
• Metal deployment (unit test), verify HPA skipped
• Verify Deployment replicas (unit test) set to nil when HPA active

Fixes #240