[FEATURE] GPUQuota status reconciler (#416 story 2)

[Defilan]

Feature Description

GPUQuota status reconciler that maintains live Status.UsedGPUCount and Status.UsedVRAMBytes by summing the InferenceServices in the quota's scope. Story 2 of #416.

Problem Statement

As an operator, I want each GPUQuota to show current utilization so I can see how close a team is to its cap before deployments start getting rejected.

Proposed Solution

Add internal/controller/gpuquota_controller.go:

• Watch GPUQuota plus InferenceService (map InferenceService events back to the matching GPUQuotas).
• Resolve scope: NamespaceRef (single namespace) or Selector (label-matched namespaces). Reject specs that set both (belt-and-suspenders; primary validation is in the webhook story).
• Aggregate GPU count and VRAM across in-scope InferenceServices; write Status.UsedGPUCount / Status.UsedVRAMBytes within one reconcile.
• Owner/finalizer handling so GPUQuota deletion cleans up cleanly.

Scope

In: status aggregation reconciler + wiring in cmd//manager setup. Out: admission decisions (that is #416-C/#416-D). This story never rejects anything; it only reports usage.

Acceptance Criteria (provable in-workspace)

• Envtest: create a GPUQuota and several InferenceServices in scope; assert UsedGPUCount/UsedVRAMBytes aggregate correctly and update when an InferenceService is added/removed.
• Envtest: deleting the GPUQuota reconciles cleanly (no leaked finalizer).
• make test and make lint pass; cross-arch lint clean.

Dependencies

• Depends on [FEATURE] GPUQuota CRD types + validation (#416 story 1) #1092 (needs the GPUQuota type). Do not start until [FEATURE] GPUQuota CRD types + validation (#416 story 1) #1092 is merged to main.

Priority

• High