[fix](load) Fix the issue of high-concurrency single-replica load get…

…ting stuck (#42297)

In high-concurrency single-replica load, the tablet_writer_add_block RPC
may occupy the _heavy_work_pool completely, causing the
response_slave_tablet_pull_rowset RPC to have no available threads for
processing. As a result, tablet_writer_add_block waits indefinitely for
a response from the slave tablet, leading to the import getting stuck
until it times out.

response_slave_tablet_pull_rowset is relatively lightweight, so it can
be handled by the _light_work_pool.

be/src/service/internal_service.cpp

@@ -1966,7 +1966,7 @@ void PInternalServiceImpl::_response_pull_slave_rowset(const std::string& remote
 void PInternalServiceImpl::response_slave_tablet_pull_rowset(
         google::protobuf::RpcController* controller, const PTabletWriteSlaveDoneRequest* request,
         PTabletWriteSlaveDoneResult* response, google::protobuf::Closure* done) {
-    bool ret = _heavy_work_pool.try_offer([txn_mgr = _engine.txn_manager(), request, response,
+    bool ret = _light_work_pool.try_offer([txn_mgr = _engine.txn_manager(), request, response,
                                            done]() {
         brpc::ClosureGuard closure_guard(done);
         VLOG_CRITICAL << "receive the result of slave replica pull rowset from slave replica. "