Remove exercise.error.checker in favor of exercise.error.check.code and -error-check chunks

[cpsievert]

Fixes #356
Fixes #425
Fixes rstudio/gradethis#140
Related rstudio/gradethis#143

After feedback from @garrettgman, and discussion with @schloerke, it's becoming apparent the exercise.error.checker probably isn't the right approach to exercise (evaluation) error checking.

This PR replaces exercise.error.checker with a slightly different option, exercise.error.check.code, which defines what checking code to run when an exercise error occurs. This option, in some sense, provides a global default value, which can be overridden for a particular exercise with a -error-check chunk. Overall this seems like a much cleaner solution than exercise.error.checker since it decouples error checking code from -code-check (pre-evaluation) and -check (post evaluation)

Testing notes

Install remotes::install_github("rstudio/learnr"); remotes::install_github("rstudio-education/gradethis"), then rmarkdown::run() the following document:

1. Enter mtcars, then click "Submit". You should see positive (green) feedback.
2. Enter cars, then click "Submit". You should see negative feedback with "This is the cars (not mtcars) dataset." in the message
3. Enter mtc, then click "Submit". You should see negative feedback with both the submission (mtc) and the solution (mtcars)
4. Enter mtc, then click "Run Code". You should see negative feedback that says object 'mtc' not found

```{r setup, include=FALSE}
library(learnr)
gradethis::gradethis_setup()
```

* Enter `mtcars`

```{r mtcars, exercise = TRUE}

```

```{r mtcars-solution}
mtcars
```

```{r mtcars-check}
grade_result(
  fail_if(~identical(.result, cars), "This is the cars (not mtcars) dataset."),
  pass_if(~identical(.result, mtcars))
)
```