Re: Using AF_UNIX sockets always for tests on Windows

Thanks Tom, Andres, Juan José, Andrew for your feedback. I agree that
it's a better "OS harmonisation" outcome if we can choose both ways on
both OSes. I will leave the 0003 patch aside for now due to lack of
time, but here's a rebase of the first two patches. Since this is
really just more cleanup-obsolete-stuff background work, I'm going to
move it to the next CF.

0001 -- Teach mkdtemp() to care about permissions on Windows.
Reviewed by Juan José, who confirmed my blind-coded understanding and
showed an alternative version but didn't suggest doing it that way.
It's a little confusing that NULL "attributes" means something
different from NULL "descriptor", so I figured I should highlight that
difference more clearly with some new comments. I guess one question
is "why should we expect the calling process to have the default
access token?"

0002 -- Use AF_UNIX for pg_regress. This one removes a couple of
hundred Windows-only lines that set up SSPI stuff, and some comments
about a signal handling hazard that -- as far as I can tell by reading
manuals -- was never really true.

0003 -- TAP testing adjustments needs some more work based on feedback
already given, not included in this revision.