We do this for loo_subsample() in the large data scenario. But we should also probably do this for standard loo() for anyway working with very large data to avoid memory issues and just compute the log-likelihood on the fly.
See arviz-devs/arviz#2197.