[xla:cpu] Optimize ThunkExecutor::Execute part #1

name                                     old cpu/op   new cpu/op   delta
BM_SelectAndScatterF32/128/process_time   889µs ± 1%   740µs ± 3%  -16.70%
BM_SelectAndScatterF32/256/process_time  3.64ms ± 2%  3.00ms ± 1%  -17.64%
BM_SelectAndScatterF32/512/process_time  15.3ms ± 1%  13.1ms ± 3%  -14.61%

PiperOrigin-RevId: 658063846

xla/service/cpu/runtime/thunk_executor.cc

@@ -162,6 +162,12 @@ tsl::AsyncValueRef<ThunkExecutor::ExecuteEvent> ThunkExecutor::Execute(
   Execute(state.get(), params, ReadyQueue(source_.begin(), source_.end()),
           /*lock=*/params.session.Join());
 
+  // If execution already completed (all kernels executed in the caller thread),
+  // immediately return the result to avoid wasteful reference counting below.
+  if (ABSL_PREDICT_TRUE(state->execute_event.IsAvailable())) {
+    return std::move(state->execute_event);
+  }
+
   // Move execute state to the execute event callback to ensure that it is kept
   // alive while thunk executor has pending tasks.
   auto execute_event = state->execute_event;