Analysis of the expected $L_2$ error of an over-parametrized deep neural network estimate learned by gradient descent without regularization