Work around NaN signbit bug on Solaris 11.4 (re: 3dfab7f)

On Solaris 11.4, when ksh is compiled with Solaris Studio cc (but
not with gcc), signbit(NAN) returns true and signbit(-NAN) returns
false! Which is of course precisely backward. This was showing up
as some regression test failures in arith.sh, as of the referenced
commit.

src/lib/libast/features/float:
- Add a test for this bug.

src/lib/libast/sfio/sfcvt.c:
- This is the function where float values are converted to strings
  for printf and the arithmetic subsystem. Compile in a workaround
  (just invert the sign again) if the bug is detected.

Author: McDutchie

src/lib/libast/features/float

@@ -1082,3 +1082,13 @@ elif	tst need_ast_pow_funs -lm note{ are IEEE pow function replacements needed }
 	#endif
 }
 endif
+
+# this bug was found on Solaris 11.4 Studio cc
+tst	nan_signbit_inverted note{ NaN signbit inverted }end execute{
+	#include <math.h>
+	int main(void)
+	{
+		double p = NAN, n = -NAN;
+		return !signbit(p) && signbit(n);
+	}
+}end

src/lib/libast/sfio/sfcvt.c

@@ -102,7 +102,9 @@ char* _sfcvt(void*	vp,		/* pointer to value to convert	*/
 
 		if (isnan(f))
 		{
-#if _lib_signbit
+#if _lib_signbit && _nan_signbit_inverted
+			if (!signbit(f))
+#elif _lib_signbit
 			if (signbit(f))
 #else
 			if (f < 0)
@@ -250,7 +252,9 @@ char* _sfcvt(void*	vp,		/* pointer to value to convert	*/
 
 		if(isnan(f))
 		{
-#if _lib_signbit
+#if _lib_signbit && _nan_signbit_inverted
+			if (!signbit(f))
+#elif _lib_signbit
 			if (signbit(f))
 #else
 			if (f < 0)