gossip: instead of refresh interval, have routing know prune_timeout.

This is twice the 'update_channel_interval' we get handed.

We delete the non-existent channel_add_connection and delete_connection
declarations from the header too.

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

tests/test_lightningd.py

@@ -1854,7 +1854,7 @@ class LightningDTests(BaseLightningDTests):
         assert [c['active'] for c in l2.rpc.listchannels()['channels']] == [True, True]
         assert [c['public'] for c in l2.rpc.listchannels()['channels']] == [True, True]
 
-    @unittest.skip("Temporarily broken for short pruning times")
+    @unittest.skipIf(not DEVELOPER, "needs DEVELOPER=1 for --dev-broadcast-interval")
     def test_gossip_pruning(self):
         """ Create channel and see it being updated in time before pruning
         """
@@ -1889,6 +1889,12 @@ class LightningDTests(BaseLightningDTests):
         ])
 
         # Now kill l3, so that l2 and l1 can prune it from their view after 10 seconds
+
+        # FIXME: This sleep() masks a real bug: that channeld sends a
+        # channel_update message (to disable the channel) with same
+        # timestamp as the last keepalive, and thus is ignored.  The minimal
+        # fix is to backdate the keepalives 1 second, but maybe we should
+        # simply have gossipd generate all updates?
         time.sleep(1)
         l3.stop()