feat(weight-transfer): support router-discovered NIXL autoscaling

[S1ro1]

Summary

PoC support for adding inference replicas to a running NIXL + ModelExpress RL job through the inference router.

• adds router_url to the shared trainer/orchestrator NIXL configuration
• moves inference-worker discovery and /init_broadcaster setup to the trainer
• discovers backend admin URLs through the router's /workers endpoint
• assigns stable rank ranges to newly discovered inference servers
• updates ModelExpress participant counts as the inference pool grows
• keeps orchestrator pause/update/resume traffic on the router endpoint

Validated topology

Validated on the H200 E2E Kubernetes cluster with:

• two trainer nodes, 16 trainer ranks, EP=8
• two initial TP=8 inference replicas
• one additional TP=8 inference replica started about 90 seconds later
• Laguna-XS.2, 100-step math run, NIXL + ModelExpress

The third replica registered as unavailable for rollout traffic, joined the next weight update as inference ranks 16–23, became routable after that update, and the job continued with 24 participating inference ranks.

Stack

• Includes the now-merged strided-view replay prerequisite: fix(nixl): replay strided views on inference #3113
• Router companion: feat: gate Kubernetes workers until weight update router#48
• 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 #3060

Current PoC limitations

• only homogeneous inference servers are supported; ranks per server are fixed from the initial pool
• scale-up was validated; arbitrary shrink/grow churn and stable-URL pod replacement still need lifecycle work
• rank assignments and membership are process-local and are not restored after trainer/router restart
• a worker failure during an in-flight synchronization remains fail-stop

Checks

• ruff format --check on changed Python files
• ruff check on changed Python files
• end-to-end staggered 2×TP8 → 3×TP8 H200 run