Modeling an empty tsibble (with key) should not error

In batch processing consistency is king. Sometimes we might apply filter on a dataset and remove all the data. Still calling model on such a tsibble should, in my opinion, not produce an error but return an empty mable. This allows us to write simple and consistent pipelines without unnecessary error checking.

This works:

> tsibble(i = 1:12, quantity = 1, index = "i") %>% filter(i > 1) %>% model(x = MEAN(quantity))
# A mable: 1 x 1
        x
  <model>
1  <MEAN>
> tsibble(i = 1:12, quantity = 1, index = "i") %>% filter(i > 12) %>% model(x = MEAN(quantity))
# A mable: 1 x 1
             x
       <model>
1 <NULL model>
Warning message:
1 error encountered for x
[1] All observations are missing, a model cannot be estimated without data.

but having a tsibble with key fails:

tsibble(i = 1:12, k = 1, quantity = 1, index = "i", key = "k") %>% filter(i > 12) %>% model(x = MEAN(quantity))
Error in `[[.default`(mdl, 1) : subscript out of bounds
>