Measuring What Matters: Connecting AI Ethics Evaluations to System Attributes, Hazards, and Harms