Automatic merge from submit-queue

Reduce kubectl calls from O(#nodes) to O(1) in cluster logdump

Ref https://github.com/kubernetes/kubernetes/issues/48513

Each node's logexporter is made to write a file to a GCS directory on success (https://github.com/kubernetes/test-infra/pull/3782).
We now use that directory as a registry of successful nodes and get it through a single "gsutil ls" call. This:
- reduces the current waiting time for logexporter in 5k-node cluster from >1hr to <10s.
- eliminates dependency on `kubectl logs` calls which seem to be unreliable sometimes (e.g when kubelet (or apiserver) is down)

/cc @kubernetes/sig-scalability-misc @wojtek-t @gmarek @fejta