Add router-visible PD Engine/Decoder pair identity

[YouNeedCryDear]

What to build

Add router-visible identity metadata for RDMA-local PD Engine/Decoder pairs. Once topology-aware placement defines the selected topology domain, OME should label or annotate Engine and Decoder pods so the router can discover matched prefill/decode pairs and avoid selecting an Engine from one rack or RDMA fabric with a Decoder from another.

The router-facing contract should keep placement and routing concerns separate: scheduling decides where pods land, while the router consumes stable pair and topology metadata to pick a local pair for prefill/decode work. This should support current service-discovery selector flows and leave room for future topology-aware routing or KV-transfer policies.

Acceptance criteria

• Engine and Decoder pods expose a stable pair identity that is scoped to the InferenceService and selected topology domain.
• Pods expose the selected topology domain key/value in router-consumable labels or annotations.
• Router discovery config can select matched Engine/Decoder pairs instead of only independent Engine and Decoder pools.
• The router can avoid cross-rack or cross-fabric prefill/decode handoff when pair metadata is present.
• Behavior is backwards compatible when pair metadata is absent or the RDMA locality policy is disabled.
• The pair identity is stable across ordinary reconciles and changes only when placement/topology changes require it.
• Tests cover label/annotation generation, router selector/config generation, missing metadata fallback, and multiple pair/topology-domain scenarios.
• Docs explain how router pairing relates to RDMA placement, what metadata is produced, and how operators can inspect the selected pairs.

Blocked by

• Add topology-aware RDMA placement for PD Engine/Decoder workloads #625