Fork per-node query action to search_coordination pool

The runtime of this action can heavily depend on the number of shards
queried, how quickly queries execute and how likely the can_match we
execute during the query phase is to return `false`.
As we shouldn't have O(n) cost tasks that may run unpredictably long,
lets fork this to search_coordination like we do for can_match and
field_caps which both have similar runtime cost.

Author: original-brownbear

server/src/main/java/org/elasticsearch/action/search/SearchQueryThenFetchAsyncAction.java

@@ -559,7 +559,7 @@ static void registerNodeSearchAction(
         final int searchPoolMax = threadPool.info(ThreadPool.Names.SEARCH).getMax();
         transportService.registerRequestHandler(
             NODE_SEARCH_ACTION_NAME,
-            EsExecutors.DIRECT_EXECUTOR_SERVICE,
+            threadPool.executor(ThreadPool.Names.SEARCH_COORDINATION),
             NodeQueryRequest::new,
             (request, channel, task) -> {
                 final CancellableTask cancellableTask = (CancellableTask) task;