Distributed Gradient Descent with Many Local Steps in Overparameterized Models