A Gram-Gauss-Newton Method Learning Overparameterized Deep Neural Networks for Regression Problems