Determine how NAs are handled

[dajmcdon]

In the canned arx_forecaster NAs need to be handled.

At train time:

• current is to omit all NAs.
• There are many potential imputation options in {recipes}, though I suspect none work for panel data OOB.

At test time:

• causes mapply() error when trying to use percent_cli as a covariate in arx_forecaster() #128
• The bug is because NAs in predictors are omitted. But this omits everything when applied to the result of get_test_data()
• On the one hand, we need some values, but what to do when there is data latency.
• Applies only to the canned forecaster.
• Could default to "locf", which seems easiest for now.

[dajmcdon]

Closed by #134