fix a problem with shrinkv tail processing

Tail processing in shrinkv had an implicit assumption of round-down, but of
course we round to nearest. Thanks angelmixu.

see https://github.com/libvips/libvips/issues/1396

ChangeLog

@@ -7,6 +7,7 @@
 - fix build with GM
 - add locks for pdfium load
 - fix build with MSVC 
+- fix a problem with shinkv tail processing [angelmixu]
 
 24/5/19 started 8.8.1
 - improve realpath() use on older libc

libvips/resample/shrinkv.c

@@ -329,12 +329,12 @@ vips_shrinkv_gen( VipsRegion *or, void *vseq,
 		r->top + r->height >= or->im->Ysize ) {
 		/* First unused scanline. resample->in->Ysize because we want
 		 * the height before the embed.
+		 *
+		 * Because we round to nearest, unused can be negative.
 		 */
 		int first = or->im->Ysize * shrink->vshrink;
 		int unused = resample->in->Ysize - first;
 
-		g_assert( unused >= 0 );
-
 		for( y = 0; y < unused; y++ ) { 
 			VipsRect s;