* Return a reasonable value from State::iterations() even before starting a benchmark

* Optimize State::iterations() for started case.