feat: Add Scale-to-Zero configuration to KServe InferenceService

Problem

The current manifests/inference-service.yaml has no autoscaling configuration. The InferenceService will keep GPU nodes running continuously even when idle, which is costly and inefficient for the bursty, on-demand query patterns of a documentation assistant.

Proposed Solution

Add Scale-to-Zero using KServe's Knative-based autoscaling annotations:

• Set minReplicas: 0 to allow full scale-down when idle
• Configure scaleTarget and scaleMetric (e.g. concurrency-based)
• Document cold-start latency tradeoff in README
• Add optional KEDA-based scaling config as an alternative for
  clusters without Knative

Why This Matters

Scale-to-Zero for bursty workloads is explicitly listed as a key deliverable in the GSoC project scope for this repo. This change also makes the reference architecture more cost-practical for community members deploying on OCI or other clouds.

References

• KServe autoscaling docs: https://kserve.github.io/website/latest/modelserving/autoscaling/autoscaling/
• GSoC proposal: Agentic RAG on Kubeflow

I'd like to work on this.

Start with manifests/inference-service.yaml and the referenced KServe autoscaling documentation to identify the repository's InferenceService configuration. Update the manifest for minReplicas, scaleTarget, and scaleMetric, document cold-start tradeoffs in README, and determine whether the optional KEDA alternative is in scope.