More precise bounds for of_float conversion to small ints (#137)

This patch improves the bounds used in of_float to decide whether the double can fit into an OCaml 63-bit tagged int.

The upper bound is 0x3ffffffffffffe00, which is exactly representable as a double (2^53-1)*2^9 = (2-2^-52) * 2^61 as it is 1.1...1 * 2^61 with 52 ones after the 1, and is less than 2^62, so it is representable as an OCaml int.   The double just after 0x3ffffffffffffe00 is 0x4000000000000000, or 2^62, which does not fit an OCaml int.

For the lower bound, -0x4000000000000000 is exactly representable as a double, but also the smallest number that fits an OCaml int.

caml_z.c

@@ -160,12 +160,10 @@ extern "C" {
 #endif
 #define Z_FITS_INT(v)  ((v) >= Z_MIN_INT && (v) <= Z_MAX_INT)
 
-/* Z_MAX_INT may not be representable exactly as a double => we use a
-   lower approximation to be safe
- */
+/* greatest/smallest double that can fit in an int */
 #ifdef ARCH_SIXTYFOUR
-#define Z_MAX_INT_FL    0x3ffffffffffff000
-#define Z_MIN_INT_FL    (-Z_MAX_INT_FL)
+#define Z_MAX_INT_FL    0x3ffffffffffffe00
+#define Z_MIN_INT_FL    (-0x4000000000000000)
 #else
 #define Z_MAX_INT_FL    Z_MAX_INT
 #define Z_MIN_INT_FL    Z_MIN_INT