Asymptotic Risk of Overparameterized Likelihood Models: Double Descent Theory for Deep Neural Networks