Improved learning theory for kernel distribution regression with two-stage sampling