Use check_outcome for all fit paths

[juliasilge]

Closes #621

This PR updates form_form() and form_xy() so that all four fit paths use the same check_outcome() helper function.

parsnip::fit(
  parsnip::decision_tree(mode = "regression"),
  status ~ wind,
  data = dplyr::storms)
#> Error in `check_outcome()` at parsnip/R/fit_helpers.R:9:4:
#> For a regression model, the outcome should be numeric.

^{Created on 2021-12-14 by the reprex package (v2.0.1)}

[hfrick]

Looks good! The test you added runs xy_form() and there are some which use form_form(). Do we also have one for form_xy()? If not, we could add one with glmnet.

library(parsnip)

linear_reg(penalty = 0.1) %>%
  set_engine("glmnet") %>%
  fit(status ~ wind, data = dplyr::storms)
#> Error: For a regression model, the outcome should be numeric.

^{Created on 2021-12-15 by the reprex package (v2.0.1)}

[juliasilge]

That would have to go into extratests because no glmnet tests get executed here in parsnip. I guess the pass through the fitting is different enough that it's worth doing that.

[github-actions]

This pull request has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.