Safety Cases: How to Justify the Safety of Advanced AI Systems