Bounding, Concentrating, and Truncating: Unifying Privacy Loss Composition for Data Analytics