With '%', non-ascii worked because the '%' automatically got promoted to
unicode.  With format that doesn't happen, which led to encoding errors.  This
fix goes back to using %, and adds a test to make sure non-ascii string values
work in iterdump.

Although there is not a regression in Python2, we make the same update here to
keep the code bases in sync.

(The fix for issue 9750 introduced a regression in Python 3 by sorting the row
objects returned by fetchall.  But if a row_factory such as sqlite3.Row is
used, the rows may not be sortable (in Python3), which leads to an exception.
The sorting is still a nice idea, so the patch moves the sort into the sql.)

Fix and test by Peter Otten.

The fix for issue 9750 introduced a regression by sorting the row objects
returned by fetchall.  But if a row_factory such as sqlite3.Row is used, the
rows may not be sortable (in Python3), which leads to an exception.  The
sorting is still a nice idea, so the patch moves the sort into the sql.

Fix and test by Peter Otten.

Closes #9750

Closes #9750