Fix tart-kubelet restart leaving VMs stuck in “already running” loop

Tart-kubelet restarts no longer strand long-lived VM workloads on the macOS fleet. Previously a restart left already-running VMs (such as the xcresult-processor) orphaned: the kubelet tried to re-run them via tart run, hit “VM is already running” every ~30s, and the Pod never reached Ready, which could wedge helm --wait --atomic production deploys. Recovery now uses a fast liveness check (IsRunning/pgrep) instead of the slow tart ip probe that a stopped golden could starve, and createPod self-heals when it encounters an already-running clone by registering it in the store instead of retrying. Result: tart-kubelet restarts during operator rolls are non-disruptive and deploys stop hanging on xcresult-processor.