Fix shutdown-metadata-related flakiness in SnapshotStressTestsIT (#109049)

We must not mark the master for shutdown since this triggers a master
failover, and we must keep all the blocks in place until the mark/unmark
sequence is complete. Also adds some logging around shutdown metadata
manipulation.

Found while investigating #108907 although this doesn't fix that issue.

Author: DaveCTurner

server/src/internalClusterTest/java/org/elasticsearch/snapshots/SnapshotStressTestsIT.java

@@ -1189,6 +1189,12 @@ private void startNodeShutdownMarker() {
 
                     final var clusterService = cluster.getCurrentMasterNodeInstance(ClusterService.class);
 
+                    if (node.nodeName.equals(clusterService.localNode().getName())) {
+                        return;
+                    }
+
+                    logger.info("--> marking [{}] for removal", node);
+
                     SubscribableListener
 
                         .<Void>newForked(
@@ -1252,12 +1258,15 @@ public void onFailure(Exception e) {
                                     @Override
                                     public void clusterStateProcessed(ClusterState initialState, ClusterState newState) {
                                         l.onResponse(null);
+                                        logger.info("--> unmarked [{}] for removal", node);
                                     }
                                 }
                             )
                         )
 
-                        .addListener(mustSucceed(ignored -> startNodeShutdownMarker()));
+                        .addListener(
+                            ActionListener.releaseAfter(mustSucceed(ignored -> startNodeShutdownMarker()), localReleasables.transfer())
+                        );
 
                     rerun = false;
                 } finally {