Check for amount of training data

[dajmcdon]

This is related to #36 and #53.

We should add a check_training_nobs() or some such name.

Desired behaviour:

• Examine the recipe. At bake time, see if we have at least nobs without NAs.
• If we don't have enough data, do something to predict() so that it outputs the right number of rows for the target as well as the right columns, but with all NA predictions. So if your frosting creates a .pred and a .pred_distn column, those would exist, but contain NAs.

A complication here is that, if everything above works, calling layer_naomit() would result in an empty epi_df. This should NOT happen.

See also #107 .

[dsweber2]

Just want to note that figuring out that this was the problem I was having took ~a day. The specific case I hit was no training obs that were ahead days ahead while also having the max of lags days behind (caused by epix_slide rounding down rather than up in the number of days it gave us, but that's a topic for somewhere else). Would definitely be useful!

[dshemetov]

Closed by #283. Can reopen if we find unaddressed edge cases.