feat(epp): hub mode for multi-cluster routing

[hexfusion]

What would you like to be added:

/feature

A "hub mode" that generalizes the EPP scheduling candidate one tier up, from pod to destination (a local InferencePool, a peer cluster, or an external provider), so the existing Filter/Score/Pick pipeline handles multi-cluster and multi-provider routing. Destinations are Endpoints in the hub's own InferencePool and the cross-cluster signal is the spoke EPP's existing pool-aggregate metrics, so there is no new scheduler and no core scheduling changes; the additions are a candidate producer and a locality scorer.

Design Doc: Hub and Spoke: multi-cluster llm-d

Why is this needed:

llm-d has no upstream path for multi-cluster and external-provider routing today. The EPP already owns the scheduling decision and the load signal within a cluster, so generalizing its candidate to a destination extends that to multi-cluster on the same pipeline, with no new component.

related work items

• feat(datalayer): extend file-discovery to support cluster endpoints #1903 file-discovery: cluster (hostname) endpoints (PR feat(datalayer): extend file-discovery to support cluster endpoints #1927), incl. EPP-side hostname->IP resolution
• feat(datalayer): Add spoke-epp engine type for cluster metrics extraction #1913 spoke-epp engine type for cluster metrics extraction
• epp: serve the metrics endpoint over mTLS #1918 serve the metrics endpoint over mTLS (PR epp: serve the metrics endpoint over mTLS #1919)
• datalayer: TLS options (CA verify + mTLS client cert) for metrics scrape #1858 TLS options (CA verify + client mTLS) for the scrape
• [Feature] Hot-reload the metrics/models scrape client cert on rotation #2045 hot-reload the scrape client cert on rotation
• epp: generalize endpoint identity #2147 epp: generalize endpoint identity
• epp: add multicluster feature gate and endpoint attribute #2191

[hexfusion]

/assign