patch from the trunk.

This stops test_socket_ssl from dying with:

    TypeError: 'NoneType' object is not callable

in urlparsre.py's urljoin() when the tests are run in this order:

    test_???
    test_importhooks
    test_socket_ssl

"test_???" can be various things, but must be there.  For
example, test_urllibnet "works" to provoke the failure.

Alas, nobody actually understands _why_ test_socket_ssl fails
then, or why this hack makes the problem go away.

Amazingly, the tests just happened to run in the right order on
the 2.4 branch on two Windows buildbot slaves today, causing them
both to fail their most recent test runs before this patch.