Init-gating GPU readiness on Kubernetes

The single highest-leverage reliability fix for edge GPU workloads: never let an inference pod schedule before the GPU is actually ready.

• kubernetes
• gpu
• edge
• reliability