even OpKernel::ComputeAync have async completion, the kernel launch / or memory copy can be still launched in the same thread, this is especially true for _Send/_Recv or _HostSend/_HostRecv.

add annotation will let use see which op had issued the memory copy.

PiperOrigin-RevId: 294723865
Change-Id: I93089ccdeef4e6beca0ae571b295c47ee07225c4

tensorflow/core/common_runtime/executor.cc

@@ -1877,7 +1877,7 @@ void ExecutorState::Process(TaggedNode tagged_node, int64 scheduled_nsec) {
         };
         nodestats::SetOpStart(stats);
         {
-          profiler::TraceMe activity(
+          profiler::AnnotatedTraceMe activity(
               [&] {
                 return op_kernel->TraceString(
                     &state->ctx, /*verbose=*/profiler::TfOpDetailsEnabled());