Deep reinforced learning enables solving rich discrete-choice life cycle models to analyze social security reforms