I wrote:
> I agree.  It seems to me that it is not python-font-lock-extend-region
> itself that is slow, but rather font-lock's processing of the area
> extended by it.  So one workaround would be to limit the number of
> lines to be extended, as in the attached patch.  If this limit is
> exceeded, the area or the entire buffer must be font-locked manually
> later.  What do you think of this idea?

The cause of the slowdown seems to be in python-info-docstring-p.  So
another option would be to improve it.  Attached is a patch that
determines that if point is in parens except for "(", it is not a
docstring.  I can't think of a case where a docstring is in parens
except for "(".  However, usually more tests should be done.