Revert "pad: Don't drop LATENCY queries with default implementation"

This reverts commit 794944f779f954375fc74a3fffcc2067bba6a3e5.

Accumulating non-live latency values generally makes no sense and often
gives invalid results with min>max
diff --git a/gst/gstpad.c b/gst/gstpad.c
index bb9e95b..da9f866 100644
--- a/gst/gstpad.c
+++ b/gst/gstpad.c
@@ -3246,7 +3246,6 @@
 /* Default latency implementation */
 typedef struct
 {
-  guint count;
   gboolean live;
   GstClockTime min, max;
 } LatencyFoldData;
@@ -3278,8 +3277,7 @@
     GST_LOG_OBJECT (pad, "got latency live:%s min:%" G_GINT64_FORMAT
         " max:%" G_GINT64_FORMAT, live ? "true" : "false", min, max);
 
-    /* FIXME : Why do we only take values into account if it's live ? */
-    if (live || fold_data->count == 0) {
+    if (live) {
       if (min > fold_data->min)
         fold_data->min = min;
 
@@ -3288,9 +3286,8 @@
       else if (max < fold_data->max)
         fold_data->max = max;
 
-      fold_data->live = live;
+      fold_data->live = TRUE;
     }
-    fold_data->count += 1;
   } else if (peer) {
     GST_DEBUG_OBJECT (pad, "latency query failed");
     g_value_set_boolean (ret, FALSE);
@@ -3321,7 +3318,6 @@
   g_value_init (&ret, G_TYPE_BOOLEAN);
 
 retry:
-  fold_data.count = 0;
   fold_data.live = FALSE;
   fold_data.min = 0;
   fold_data.max = GST_CLOCK_TIME_NONE;