When I renamed `line` to `trimmed`, I forgot one.

(Previously, I had rebound `line`, and I think that made
these sorts of bugs impossible.)

This regression is covered by new tests for Markdown supported
in a subsequent commit.

This commit slots in support for formatting Markdown fenced code
blocks[1]. With the refactoring done for reStructuredText previously,
this ended up being pretty easy to add. Markdown code blocks are also
quite a bit easier to parse and recognize correctly.

One point of contention in #8860 is whether to assume that unlabeled
Markdown code fences are Python or not by default. In this PR, we make
such an assumption. This follows what `rustdoc` does. The mitigation
here is that if an unlabeled code block isn't Python, then it probably
won't parse as Python. And we'll end up skipping it. So in the vast
majority of cases, the worst thing that can happen is a little bit of
wasted work.

Closes #8860

[1]: https://spec.commonmark.org/0.30/#fenced-code-blocks