Autocompletion for references happens on the frontend. Those references
are turned into actual references on the backend, but only after
Markdown processing has happened. That means that if a reference
contains a character that Markdown might consume, it won't render
correctly. So we need to do some escaping on the frontend.

We have these potential problem characters:
https://docs.gitlab.com/ee/user/markdown.html#emphasis

1. ~ - this is ~~strikethrough~~, but only when doubled.
2. _ - used for _emphasis_, doubled is __bold__.
3. * - also used for *emphasis*, doubled is **bold** also.
4. ` - used for `code spans`, any number works.

We don't need to escape `-` any more. When it comes to being inside a
word:

1. a~~b~~ has strikethrough, so it needs to be escaped everywhere.
2. a_b_ has no emphasis (see [a]) so it only needs to be escaped at the
   start and end of words.
3. a*b* has emphasis, so it needs to be escaped everywhere.
4. a`b` has a code span, so it needs to be escaped everywhere.

Or, in code terms:

1. Always escape ~~, *, and ` when being inserted by autocomplete.
2. Escape _ when it's either at the beginning or the end of a word.

[a]: https://docs.gitlab.com/ee/user/markdown.html#multiple-underscores-in-words