Before, any elements that were siblings of matching text nodes got
destroyed, replaced with their normalised `textContent`. That is, given
this DOM:

```html
<p>
  Foo <i>bar</i> qux.
</p>
```

If you searched for `foo`, then the above would be transformed into
something like:

```html
<p>
  <mark>Foo</mark> bar qux.
</p>

<!-- But the correct DOM would be: -->
<p>
  <mark>Foo</mark> <i>bar</i> qux.
</p>
```

And then if you cleared the search, it would become:

```html
<p>
  Foo bar qux.
</p>

<!-- But the correct DOM would be the same as before the search: -->
<p>
  Foo <i>bar</i> qux.
</p>
```

The crux of the problem was that the *highlighting* algorithm wasn't the
inverse of the *highlight clearing* algorithm. Now, it is, and only text
nodes are modified. Matching text nodes are split into non-matching text
nodes and `mark` elements containing the matched search text, and that
collection of nodes replaces the original text node.

Addresses https://gitlab.com/gitlab-org/gitlab/-/issues/350494.

Changelog: fixed