Fix order of -I switches for building pg_regress.o.

We need the -I switch for libpq_srcdir to come before any -I switches
injected by configure.  Otherwise there is a risk of pulling in a
mismatched version of libpq_fe.h from someplace like
/usr/local/include, if the platform has another Postgres version
installed there.  This evidently accounts for today's buildfarm
failures on "anaconda".

In principle the -I switch for src/port/ is at similar hazard, and has
been for a very long time.  But the only .h files we keep there are
pg_config_paths.h and pthread-win32.h, neither of which get installed
on Unix-ish systems, so the odds of picking up a conflicting header
seem pretty small.  That doubtless accounts for the lack of prior
reports.

Back-patch to v17 where pg_regress acquired a build dependency on
libpq_fe.h.  We could go back further to fix the hazard for src/port/
in older branches, but it seems unlikely to be worth troubling over.

Reported-by: Nathan Bossart <nathandbossart@gmail.com>
Author: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/Z-MhRzoc7t-nPUQG@nathan
Backpatch-through: 17

src/test/regress/GNUmakefile

@@ -40,7 +40,7 @@ pg_regress$(X): pg_regress.o pg_regress_main.o $(WIN32RES) | submake-libpgport
 
 # dependencies ensure that path changes propagate
 pg_regress.o: pg_regress.c $(top_builddir)/src/port/pg_config_paths.h
-pg_regress.o: override CPPFLAGS += -I$(top_builddir)/src/port -I$(libpq_srcdir) $(EXTRADEFS)
+pg_regress.o: override CPPFLAGS := -I$(top_builddir)/src/port -I$(libpq_srcdir) $(EXTRADEFS) $(CPPFLAGS)
 
 # note: because of the submake dependency, this rule's action is really a no-op
 $(top_builddir)/src/port/pg_config_paths.h: | submake-libpgport