Improve errors for set_mode() and set_engine()

[juliasilge]

Closes #426

This PR makes our handling of a missing argument and a bad argument a bit more consistent for set_mode() and set_engine(). There was some handling coded up in both cases but we weren't really getting those benefits for missing (or even NULL) inputs before. This rearranging should make help things out.

The already existing warning/replacement for engines now works:

library(parsnip)
linear_reg() %>% set_engine()
#> Warning: `engine` was NULL and updated to be `lm`
#> Linear Regression Model Specification (regression)
#> 
#> Computational engine: lm

## same as before
linear_reg() %>% set_engine("ranger")
#> Error: Engine 'ranger' is not available. Please use one of: 'lm', 'glmnet', 'stan', 'spark', 'keras'

^{Created on 2021-05-11 by the reprex package (v2.0.0)}

For mode, I just set this up to tell the user their options:

library(parsnip)
rand_forest() %>% set_mode()
#> Error: `mode` should be one of 'classification', 'regression', 'censored regression'
rand_forest() %>% set_mode(NULL)
#> Error: `mode` should be one of 'classification', 'regression', 'censored regression'
rand_forest() %>% set_mode("ranger")
#> Error: `mode` should be one of 'classification', 'regression', 'censored regression'

^{Created on 2021-05-11 by the reprex package (v2.0.0)}

I did need to adjust some tests, switching between errors and warnings, but I don't think any of that is substantive.

[juliasilge]

For both set_mode() and set_engine(), we could set a default value of NULL for the mode/engine and remove those lines like:

if (rlang::is_missing(engine)) engine <- NULL

Do we want to have a default value show up in the docs? It seems a little weird when a value of NULL just triggers these warnings / errors.

[topepo]

There might be a number of places in the code and docs to remove traces of "unknown" mode. TBH I'm fine with leaving that possible mode but not referencing it in warnings or errors that happen during checks about model fitting.

[juliasilge]

I looked at this again today and changed my mind about the set_mode() approach. Now it all goes through check_spec_mode_val() and there is no more returning the generic parsnip:::all_modes in some situations:

library(parsnip)
logistic_reg() %>% set_mode()
#> Error: `mode` should be one of: 'unknown', 'classification'
logistic_reg() %>% set_mode("regression")
#> Error: `mode` should be one of: 'unknown', 'classification'

rand_forest() %>% set_mode()
#> Error: `mode` should be one of: 'unknown', 'classification', 'regression'
rand_forest() %>% set_mode(NULL)
#> Error: `mode` should be one of: 'unknown', 'classification', 'regression'
rand_forest() %>% set_mode("ranger")
#> Error: `mode` should be one of: 'unknown', 'classification', 'regression'

^{Created on 2021-05-12 by the reprex package (v2.0.0)}

I tend to think it is OK to leave "unknown" for now and we can make a call about that in the future.

[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.