Summary

• Sets deployment strategy to Recreate for GPU workloads (gpuCount > 0) to prevent scheduling deadlock during rolling updates
• Non-GPU workloads continue using the Kubernetes default (RollingUpdate)
• Adds test assertions for both GPU and CPU-only strategy behavior

Problem

When all GPUs on a node are occupied, RollingUpdate creates a deadlock: the new pod cannot schedule without a GPU, and the old pod won't terminate until the new pod is Ready. This is especially common in homelab/small clusters with no spare GPUs.

How it works

The fix adds a Recreate strategy assignment inside the existing if gpuCount > 0 block in constructDeployment(), alongside the GPU toleration logic. This means the old pod terminates first, freeing the GPU for the replacement pod. Brief downtime during updates is acceptable — the alternative is a permanent deadlock.

Existing GPU InferenceService deployments will pick up the fix on the next reconcile cycle.

Fixes #192

Test plan

• make test passes — all existing + new tests
• Verify GPU deployment has strategy: Recreate via kubectl get deploy <name> -o yaml
• Verify CPU-only deployment retains default RollingUpdate
• Trigger a GPU InferenceService update and confirm old pod terminates before new pod creates