Counterexample-Guided Strategy Improvement for POMDPs Using Recurrent Neural Networks