Fix race in the VSL segment counter initialization

Fix a stupid misstake where the segment counter in the VSL segment was
set just before the entire struct was cleared during
initialization. This didn't cause problems in the common case because
it would be set correctly on the very first log line produced. But it
did open a race where an unfortunate early log client would assume it
was lagging indefinetely behind varnishd on the log reading and report
overflow error. This was most commonly experienced with varnishtest.

Fixes: #2241

bin/varnishd/cache/cache_shmlog.c

@@ -470,13 +470,13 @@ VSM_Init(void)
 	   problems with regard to readers on that event visible */
 	vsl_segment_n = UINT_MAX - VSL_SEGMENTS + 1;
 	AZ(vsl_segment_n % VSL_SEGMENTS);
-	vsl_head->segment_n = vsl_segment_n;
 	vsl_ptr = vsl_head->log;
 	*vsl_ptr = VSL_ENDMARKER;
 
 	memset(vsl_head, 0, sizeof *vsl_head);
 	vsl_head->segsize = vsl_segsize;
 	vsl_head->offset[0] = 0;
+	vsl_head->segment_n = vsl_segment_n;
 	for (i = 1; i < VSL_SEGMENTS; i++)
 		vsl_head->offset[i] = -1;
 	VWMB();