fix: retry stale GPU prewarm pods

[bradhilton]

Summary

• retry steady-state GPU prewarm DaemonSet pods whose mutable latest tag resolves to a stale digest
• keep the digest verification as the gate before the workflow smoke-launches SkyPilot

Testing

• bash -n scripts/build-gpu-image.sh
• scripts/build-gpu-image.sh --help
• git diff --check

The previous main run reached the steady DaemonSet with 8/10 pods on the new digest and 2/10 pods still resolving latest to the prior digest. This patch deletes only those stale DaemonSet pods and lets Kubernetes recreate them until every refresh-tag init container resolves to the pushed digest.