Refactor layer_naomit()

[dajmcdon]

Refactors layer_naomit() to make sure that it returns NAs for the targets but gets rid of the rest. As I think about it, we actually shouldn't be dropping NAs in the predictions per se, but rather only returning the rows that have meaningful predictions (even if these are NA).

So far, it's been coincidental that the rows with NA are not meaningful. But this isn't foolproof. It may be that NAs are meaningful (as in #106) but it's also possible that they aren't (often as a result of lagging the predictors).