It seems like the markdown rendering plugins get confused by nested code
fences so we replace embedded fences with ` ` `

Instead of disabling embedded markdown rendering, we can use a longer
outer codefence.

It did require regenerating all of the test data.

Instead of just using the filename extension, use the neovim filetype
(with some specific overrides). Fall back to the filename extension.

This means we render our markdown sections with `markdown` as the type
instead of `md` and that enables rendering embedded codefences
correctly.

Hopefully the last fix on this :)

Seems like nightly returns text as the type for .txt files which breaks
our tests

This was a subtle one and it only shows up when we're not collapsing
events so it wasn't caught by the unit tests. I've since modified the
unit tests to also run the replays without collapsing to catch bugs like
this in the future.

If the first version of a part didn't result in something being
rendered, future updates to that part would also not render anything.
This was particularly likely to happen with tools as the first version
of the part is often `step-start`. Because we didn't think this was a
new part, we called `_replace_part_in_buffer` but that exits quickly
because the part hadn't been rendered anywhere.

While we could try to call `_insert_part_to_buffer` in
`_replace_part_in_buffer` if the part hadn't been rendered, that would
miss some subtle cases around error handling so the correct fix is
consider the part new in `on_part_updated` if we haven't rendered it
before.