Switch to loo.matrix() instead of loo.function() by default?

[jgabry]

I think loo.matrix() is quite a bit faster than loo.function() and the latter is really only needed when N is very large for memory issues. Super large N is not uncommon but probably the minority of rstanarm use cases, so we're not really optimizing this. brms uses loo.matrix with an option to switch to loo.function. Perhaps we should do the same?