Fix two places that thought Windows64 is indicated by WIN64 macro.

Everyplace else thinks it's _WIN64, so make these places fall in line.

The pg_regress.c usage is not going to result in any change in behavior,
only suppressing (or not) a compiler warning about downcasting HANDLEs.
So there seems no need for back-patching there.

The libpq/win32.mak usage might represent an actual bug, if anyone were
using this script to build for Windows64, which perhaps nobody is.
Given the lack of field complaints, no back-patch here either.

pg_regress.c problem found by Christian Ullrich, the other by me.

src/interfaces/libpq/win32.mak

@@ -16,7 +16,7 @@ CPU=i386
 !MESSAGE Building the Win32 static library...
 !MESSAGE
 !ELSEIF ("$(CPU)" == "IA64")||("$(CPU)" == "AMD64")
-ADD_DEFINES=/D "WIN64" /Wp64 /GS
+ADD_DEFINES=/D "_WIN64" /Wp64 /GS
 ADD_SECLIB=bufferoverflowU.lib
 !MESSAGE Building the Win64 static library...
 !MESSAGE

src/test/regress/pg_regress.c

@@ -2386,7 +2386,7 @@ regression_main(int argc, char *argv[], init_function ifunc, test_function tfunc
 
 		postmaster_running = true;
 
-#ifdef WIN64
+#ifdef _WIN64
 /* need a series of two casts to convert HANDLE without compiler warning */
 #define ULONGPID(x) (unsigned long) (unsigned long long) (x)
 #else