Re-implement scheduling-throughput measurement

Currently, we have a very crude logic for computing [scheduling-throughput](https://github.com/kubernetes/perf-tests/blob/master/clusterloader2/pkg/measurement/common/scheduling_throughput.go). We list all pods every 5s and see how many new pod very scheduled in that time. It doesn't work well in small tests where number of pods is low as we only get 1-2 representative 5s windows.

We should rewrite it to make it more reliable and accurate, e.g. use prometheus query to monitor rate of pod bindings

Start with clusterloader2/pkg/measurement/common/scheduling_throughput.go and trace the current 5-second pod-listing calculation. Review how a Prometheus query could provide scheduling-throughput data, then define coverage for small tests; done means the measurement is more reliable and accurate than the current sparse-window approach.

Written by the indexing model from the issue text.