lfo returns NA if L is larger than the length of series for some group(s)

[helske]

Consider a case where T is the sample size of the largest group, L<T is the length of initial fit in lfo, and where there is less than L observations in some groups. This leads to NA elpd for those groups, which is then propagated to the final ELPD estimate. This probably happens also in a case where some series have NA values at the beginning of the series for K > L time points. It's probably enough to add na.rm = TRUE to few places in the end to handle both of these cases.

[helske]

Actually it seems that issue happens only when refitting the model, adding na.omit when storing elpds (which I think we had before) should do the trick.