Optimal Rerandomization via a Criterion that Provides Insurance Against Failed Experiments