Support additional languages for random phrases

[gadenbuie]

Fixes #534

Random phrases are now stored in data-raw/i18n_random-phrases.yml.

praise:
  en:
    - Absolutely fabulous!
    - Amazing! 
  <language key>:
    - <phrase>
encouragement:
  en:
    - Please try again.
    - Give it another try.
  <language key>:
    - <phrase>

New phrases in other languages (not direct translations of the English phrases) can be added to that YAML file. Maintainers will need to run data-raw/i18n_translations.R to update the internal data.

I added a language argument to random_praise() and random_encouragement() that can be used to set the language of the phrase. I don't recommend that argument be used other than for testing. Instead, authors can rely on these functions automatically determining which language is currently in use in the tutorial using the existing language option in the tutorial YAML header. (So language keys in i18n_random-phrases.yml need to match those in i18n_translations.yml.)

The primary challenge with this PR is that the language of the tutorial needs to be communicated from the app to the exercise evaluation, since technically the language can change even during runtime. To address this:

1. Localization in the browser now reports the current language of the tutorial back to Shiny.
2. learnr stores the current language in the session object (via write_request())
3. When an exercise is submitted, the current language is added to the exercise object as exercise$tutorial$language (via read_request())
4. evaluate_exercise() finds this language or the default language value and sets a knitr option for the exercise evaluation
5. When rendering the exercise, random_phrases() (the helper behind praise/encouragement) will look for the knitr option when picking a language, falling back to English if not set or if there aren't random phrases in the current language.

I added a helper random_phrases_add() that lets users declare random phrases for use in the tutorial outside of contributing them directly to learnr. The process is additive, so phrases are added rather than overwriting existing phrases for a given language.

The mechanism of communication here is to use shiny prerendered chunks if random_phrases_add() is called in a standard R chunk that calls an internal function to set the exercise.random_phrases chunk option. If called in the global setup chunk, then the internal function is called directly (without the prerendered chunk). The end effect is that all of the exercises will pick up the random phrases chunk option setting, which ensures that these random phrases are available when evaluating user code.

Getting the above to work required a small change to the exercise rendering. Rather than writing the exercise chunk options into the user code chunk, we now rely on chunk options being set globally in when rmarkdown::render() is called. This skips writing R objects into character strings, which was breaking when trying to embed a large list into these options since chunk headers can't be multiline. We already have the options as an R object so we just use it to directly set the chunk options via knitr_options in rmarkdown::output_format().

Finally, I bumped the current exercise version to 3, since we've now added the expectation that tutorial$language is present. (I'm not totally certain how necessary the version bump is and could be convinced to roll that back; random_phrases() will fall back to "en" if that field isn't present.)