A split-and-conquer variable selection approach for high-dimensional general semiparametric models with massive data