Nothing stabilizes or rolls back unstable writes of a removed pool, its
PGs are never peered again. The old store can't trim its journal past
them, so the journal fills up and writes to all other pools on the same
OSDs fail with -EAGAIN forever, restarts don't help.

Add a synchronous get_unstable_objects() which walks dirty_db of the old
store, and roll back unstable writes of objects outside of existing pools
with it: on start, and when a pool and its PG configuration are removed
and local PGs of the pool have stopped. The new store keeps such writes
in its metadata without harm, so it returns nothing.

@@ -505,6 +505,10 @@ void osd_t::on_change_etcd_state_hook(std::map<std::string, etcd_kv_t> & changes
         }
         apply_pg_config();
     }
+    if (pools || changes.find(st_cli->etcd_prefix+"/pg/config") != changes.end())

@@ -728,6 +732,7 @@ void osd_t::on_load_pgs_hook(bool success)
             apply_pg_count();
             apply_pg_config();
         }
+        rollback_removed_pools();

@@ -476,0 +482,4 @@
+{
+    for (auto & pc: st_cli->pool_config)
+    {
+        if (pc.second.exists)

@@ -476,0 +486,4 @@
+        {
+            continue;
+        }
+        // Wait until PGs are removed too, so that their primaries stop sending writes

@@ -476,0 +506,4 @@
+
+void osd_t::rollback_removed_pools()
+{
+    if (!bs)

@@ -476,0 +515,4 @@
+    {
+        return;
+    }
+    for (auto & pp: pgs)