When deploying to a cluster deployed on three different AZ, there is 1/3 of chance of things working

[mbaldessari]

Scenario:

1. Openshift cluster with nodes spread across three AZs (e.g. 1a, 1b, 1c)
2. Run make create-gpu-machineset it will pick the AZ of the first machineset it finds (e.g. 1a)
3. Run ./pattern.sh make install and the model-download job run on another AZ (e.g. 1c) and the pvc+pv are bound to that AZ
4. The inference service which runs on the GPU nodes (so 1a) will never be scheduled because the PVC is bound to 1c