Convergence of a projected gradient method with trust region for nonlinear constrained optimization