Switch flt_mantissa fields from long to uint32_t.

This seems to fix an error when flt_arith converts a literal
double-precision float to IEEE format.  For example, 0.5 and 0.75 got
converted to slightly below their correct values.

My host gcc for amd64 has 64-bit long, but flt_arith needs only 32
bits.  The code (at least flt_add.c) can make 32-bit overflows.  Such
overflows would set the higher bits of a 64-bit long, which might
cause problems later.

I need to use uint32_t and not int32_t because the code still uses
long, and the sign extension from int32_t to long would cause
problems.  The mantissa represents a value in [0, 2) that can't be
negative, so unsigned type is better.  Also, signed overflow is
undefined behavior in C, so flt_add.c better make overflows with
uint32_t and not int32_t.

This commit doesn't touch lang/cem/libcc.ansi/stdlib/ext_fmt.h which
continues to use unsigned long for its mantissa fields.

modules/src/flt_arith/flt_arith.h

@@ -7,13 +7,15 @@
 #ifndef __FLT_INCLUDED__
 #define __FLT_INCLUDED__
 
+#include <stdint.h>
+
 #ifndef arith
 #define arith long
 #endif
 
 struct flt_mantissa {
-	long	flt_h_32;	/* high order 32 bits of mantissa */
-	long	flt_l_32;	/* low order 32 bits of mantissa */
+	uint32_t flt_h_32;	/* high order 32 bits of mantissa */
+	uint32_t flt_l_32;	/* low order 32 bits of mantissa */
 };
 
 typedef struct flt_arith {