PG_SETMASK() archeology

Hi,

This is a follow-up for commit c94ae9d8.  It's in the spirit of other
recent changes to remove noise from ancient pre-standard systems.

The reason we introduced PG_SETMASK() in the first place was to
support one particular system that was very slow to adopt the POSIX
signals stuff:  NeXTSTEP 3.x.

From some time in the dark age before our current repo begins until
'97 we used sigprocmask() freely.  Then commit a5494a2d added a
sigsetmask() fallback for NeXTSTEP (that's a pre-standard function
inherited from '80s BSD).  In 1999 we added the PG_SETMASK() macro to
avoid repeating #ifdef HAVE_SIGPROCMASK to select between them at each
call site (commit 47937403676).  I have no personal knowledge of those
systems; I wonder if it was already effectively quite defunct while we
were adding the macro, but I dunno (NS 4.x never shipped?, but its
living descendent OSX had already shipped that year).

Then we invented a bogus reason to need the macro for a couple more
decades: our Windows simulated signal layer accidentally implemented
the old BSD interface instead of the standard one, as complained about
in commit a65e0864.

That's all ancient history now, and I think we might as well drop the
macro to make our source a tiny bit less weird for new players, with a
slightly richer interface.  Trivial patch attached.