Add a metrics to record congestion backoff counts.

When we enable congestion backoff, we should better know how many times datanodes have told client backoff.  This metrics can help us know better about congestion function.