All file-based tests are now defined as unittest test cases via a
metaclass which walks a directory and builds a unittest for each pair
of test files.

To run the tests just run `python -m unittest discover tests`. Or use
tox as the tox config has been updated to run the new tests and all nose
specific code has been removed.

The test generator tools have been removed as well. If any changes or
additions need to be made to tests, they should be implemented using
the new framework rather than with the file-based tests. Eventually,
only the PHP and pl tests should remain as file-based tests.

The current implementation was wrong as it also percent encoded query strings
(which should be plus encoded) and calling urllib.quote on the path (and
urllib.quote_plus on the query string) assumes the url is not already encoded.
What if the document author pasted a url that was already encoded? She probably
did not intend for `%20` to become `%2520`. Or did she? It is now clear to me
why many implementation do nothing to urls. Just pass them though as-is. To bad
if they are not valid HTML. HTML authors have to encodee their own urls, so I
guess markdown authors have to as well.

Also refactored the reference preprocessor to make this a little easier to
implement. Regex does more now.

Markdown markup inside angle bracktes now gets rendered properly
in all cases except when safe_mode='escape'. Also added tests.