fix(function-autoscaler): re-evaluate metric source after first LLM request

Describe the bug

In autoscaler 1.21.0, a running LLM function with no prior requests has no worker or per-function gateway series. The autoscaler selects ControlPlane and caches it for one hour, so later LLM traffic remains invisible and the function cannot scale correctly.

Steps or code to reproduce bug

1. Deploy a new LLM function with minInstances: 1.
2. Wait 2–3 minutes without sending requests.
3. Send sustained LLM traffic.
4. Observe that the autoscaler continues using ControlPlane metrics and reports 0% utilization.

Expected behavior

Absence-based fallback is re-evaluated after gateway metrics appear, and missing utilization data cannot trigger scale-to-zero.

Additional context

• Parent tracking issue: https://github.com/NVIDIA/nvcf/issues/15
• Add coverage for traffic beginning after initial metric-source selection.

By submitting this issue, you agree to follow our code of conduct and our contributing guidelines.

Start by tracing the function autoscaler’s metric-source selection and its one-hour cache, then inspect the existing autoscaler coverage around traffic beginning after initial selection. Reproduce the no-request-then-LLM-traffic sequence and verify that gateway metrics are reconsidered and missing utilization cannot cause scale-to-zero.