feat: gate Kubernetes workers until weight update

[S1ro1]

Summary

Router-side PoC for safe inference autoscaling during NIXL + ModelExpress RL runs.

• fans /pause, /update_weights, and /resume out to backend admin servers
• adds a separate usable worker state in addition to health
• registers Kubernetes-discovered workers as healthy but unusable
• excludes unusable workers from every routing policy, readiness, and health availability
• marks a worker usable only after its first successful /update_weights
• exposes is_usable from /workers for Prime-RL's transfer coordinator
• reconciles static DNS endpoints with Kubernetes-discovered pod-IP endpoints
• processes discovery events concurrently and retries cleanly after watcher errors

Stack

• Prime-RL companion: feat(weight-transfer): support router-discovered NIXL autoscaling prime-rl#3119
• Kubernetes/FFT companion and runnable example: https://github.com/PrimeIntellect-ai/fft/pull/27
• Full Jannik handoff: https://github.com/PrimeIntellect-ai/fft/blob/exp/router-autoscaling-poc/charts/prime-rl-fft/examples/NIXL_AUTOSCALING_HANDOFF.md
• Builds on merged NIXL + ModelExpress support: feat(weight-transfer): add NIXL + ModelExpress synchronization prime-rl#3060

Current PoC limitations

• usability is in-memory and is lost when the router restarts
• replacement pods that reuse a static worker URL need a stronger identity/re-registration model
• fanout is fail-stop: any backend admin failure fails the synchronization
• this is readiness gating, not request-level retry or transparent recovery for requests already executing on a killed worker

Checks

• cargo check
• complete cargo test suite: all tests passed, one pre-existing ignored test
• end-to-end staggered 2×TP8 → 3×TP8 H200 run