PR #24388 changed the markdown parsing to copy parsed text in order to
handle markdown escaping, removing the optimization to instead reuse
text from the input.

Another issue with that change was that handling of finding links within
`Text` intermixed use of `text` and `parsed`, relying on the offsets
matching up (which I believe was true in practice).

The solution is to distinguish pulldown_cmark `Text` nodes that share
bytes with the input and those that do not.

Release Notes:

- N/A