Finite-sample refinement of GMM approach to nonlinear models under heteroskedasticity of unknown form