have cluster-machines-ready wait for last node to be drained/deleted
In #4262 (closed) we observe a case where, after cluster-machines-ready finishes, lots of units start to reconcile, but this happens in parallel with the drain/delete of the last node (this is a capo run with maxSurge > 0).
The fact that the node rolling update isn't actually fully finishes when cluster-machines-ready becomes Ready, is not really intentional; I think most of us have in mind that cluster-machines-ready being ready is a point at which we consider that the set of Ready nodes is stable. This is in fact not the case.
Sometimes (and this is the case of the run analyzed in #4262 (closed)), the fact that the last node being drained is a CP node makes things worse, because the etcd member removal seems to have a side effect on kustomize-controller leases.
I think that we need to improve cluster-machines-ready to wait for the last Node to be drained and deleted.