Stop changing number of requests after handoff to MPI and don't let waiting drop to zero when we are about to requeue testsome

There is a lock, that should be necessary if we switch stealing back on, don't want to call MPI_Testsome from multiple threads