Marginal likelihood and model selection for Gaussian latent tree and forest models