Estimating linear functionals in nonlinear regression with responses missing at random