fix pixel addressing on 32-bit platforms

there was a size_t where there should have been a guint64, causing
addressing problems on very large images on 32-bit platforms (like
win32)

looks like this had been broken since March 2011

ChangeLog

@@ -2,6 +2,8 @@
 - lazy read from tiled tiff from layers other than 0 was broken
 - optional args to vips_call*() do not work, disabled (fixed correctly in
   master)
+- address calculations in files over 4gb were broken on 32-bit platforms
+  (broken since March 2011, oops)
 
 12/10/11 started 7.26.6
 - NOCACHE was not being set correctly on OS X causing performance 

libvips/include/vips/image.h

@@ -407,7 +407,7 @@ typedef struct _VipsImageClass {
 
 GType vips_image_get_type( void );
 
-extern const size_t vips__image_sizeof_bandformat[];
+extern const guint64 vips__image_sizeof_bandformat[];
 
 /* Pixel address calculation macros.
  */

libvips/iofuncs/header.c

@@ -189,8 +189,12 @@ static HeaderField old_double_field[] = {
 
 /* This is used by (eg.) VIPS_IMAGE_SIZEOF_ELEMENT() to calculate object
  * size.
+ *
+ * We can't use size_t or off_t because we have to work on some platforms
+ * which have these as 32-bit and we need to calculate addresses in large
+ * files. 
  */
-const size_t vips__image_sizeof_bandformat[] = {
+const guint64 vips__image_sizeof_bandformat[] = {
 	sizeof( unsigned char ), 	/* VIPS_FORMAT_UCHAR */
 	sizeof( signed char ), 		/* VIPS_FORMAT_CHAR */
 	sizeof( unsigned short ), 	/* VIPS_FORMAT_USHORT */