feat(dynamo): add Kubernetes DGD support

[jthomson04]

Summary

• stack the Kubernetes-specific Dynamo work from feat: Dynamo K8s Integration NVIDIA-NeMo/RL#2990 on top of the managed Dynamo/Slurm implementation in feat(dynamo): add managed Slurm generation with W&B telemetry NVIDIA-NeMo/RL#3391
• add nrl-k8s lifecycle support for suspended RayJobs, DynamoGraphDeployments, readiness, ownership, reparenting, and cleanup
• add a strict external-DGD runtime that discovers and freezes the vLLM worker fleet while leaving serving lifecycle ownership to Kubernetes
• allocate all Ray GPUs to training for external Dynamo and support collective weight synchronization without an inference Ray cluster
• carry forward the V1, V2, V3, and V5 GB300 Kubernetes recipes and document the ownership/refit model

Validation

• 275 nrl-k8s unit tests passed
• 142 focused Dynamo and weight-sync unit tests passed on the latest PR 3391 head
• Pyrefly: 0 errors across modified core modules
• Ruff lint and format checks passed
• all four Dynamo recipe/infra pairs pass nrl-k8s check
• all four Dynamo recipe/infra pairs pass RayJob dry-run rendering
• live V1 brought up the DGD and Ray resources and reached NeMo-RL model initialization; that smoke test exposed and drove fixes for strict recipe fields and external-Dynamo weight-sync construction

The live V1 two-step completion and V3 smoke remain to be captured. The validation runner lost namespace RBAC overnight, so this remains a draft until those runs can be observed end to end.

Stack

• base: feat(dynamo): add managed Slurm generation with W&B telemetry NVIDIA-NeMo/RL#3391
• Kubernetes source: feat: Dynamo K8s Integration NVIDIA-NeMo/RL#2990