Scalable Marginal Likelihood Estimation for Model Selection in Deep Learning