Residuals from non-lm engines

[dajmcdon]

Looking at the sliding.Rmd vignette, the random forest prediction intervals are very narrow. Is there a bug somewhere in the residual_quantiles()?

[dajmcdon]

Update: seemingly two bugs.

1. There's a bug in the residual calculation when residuals.fit_engine() is unavailable (it tries to use the test data).
2. We need to impute missing values. For now, I'm adding a step_impute_locf() function (though one could use recipes::step_impute_roll() if desired.