Allow grade_this_code() to provide pass-through-filter grading

[gadenbuie]

This PR relaxes the constraint applied in gradethis_exercise_checker() that gradethis grading functions must return either passing or failing feedback. This builds on rstudio/learnr#610 and uses the new stage argument set in the checker function when evaluating exercises. The constraint is now only applied to stage == "check" since this defines the terminal check action and other grading chunks might pass through.

Then, we build on the above feature by adding an action parameter to grade_this_code(), allowing the grading author to choose if grade_this_code() will pass or fail the user's code. The default is "both" to be consistent with current behavior. If either "pass" or "fail" are chosen, feedback that would result in the opposite action is ignored and no feedback is returned.

The end result is that authors can write check chunks that use static code analysis to filter out problematic results that might be computationally intensive, but that defer to the -check chunk when the code matches.

Use the correct join function to combine these
two giant tables without crashing your session.

```{r join, exercise = TRUE}
dplyr::____(giant, huge)
```

```{r join-solution}
dplyr::inner_join(giant, huge)
```

```{r join-code-check}
grade_this_code(incorrect = "Not that one", action = "fail")
```

```{r join-check}
grade_this({
  pass_if_equal(message = "Good choice!")
  fail()
})
```

If the user picks left_join(), static analysis will return a feedback message highlighting the user's mistake. If the user picks inner_join(), grade_this_code() defers to the -check chunk where the user's evaluated code is graded. (Totally made up example, but similar issues arise in plotting and many other places.)

[rossellhayes]

This is a really nice change!