Rendered text file attachments (such as Markdown) are rendered as

Unicode, for output to the client. These were getting passed directly to

the differ, which itself takes byte strings. On Python 2, this all

worked out fine. On Python 3, this blew up pretty early, but in a way

that was hard to diagnose due any errors coming from subclasses just

being logged (with a massively-huge log message) and then the exception

ignored.
The Markdown review UI specifically had another issue where we were

passing in a file handle containing byte strings, which was then getting

written back as byte strings but to a StringIO expecting Unicode

strings. While the core issue lives in Djblets (and there's another

patch up that aims to fix this), it's better if we just read the file

content and convert to Unicode first, since the Markdown library doesn't

stream the file pointer anyway (it just grabs all the data at once).
This change most of these issues by doing the following:
1) Ensures we're always passing the right string types around (both to

   Markdown and the diff chunk generators).
2) Switches the Markdown rendering call to render_markdown(), passing

   it a Unicode string instead of a file handle.
3) Adds new unit tests covering all this.
It does not address the overzealous exception sandboxing. That will be

taken care of separately.

Unit tests passed.
Verified I could view Markdown files in both rendered and source mode,

with diffs.