Created by: sneaxiy

AsyncSSAGraphExecutor is not compatible with ordered DataLoader, because:

• AsyncSSAGraphExecutor would call BuildStrategy::Apply methods for each sub-graph, so that we cannot get device index using ComputationOpHandle::GetScopeIdx().
• The ordered non-iterable DataLoader would put create_py_reader op inside main program, so that the global scope would have the persistable var of blocking queue. Inside AsyncSSAGraphExecutor, we have to init variables in local scopes in the reverse order, so that each create_py_reader op of each device would get the right blocking queue. Otherwise, all create_py_reader ops in all devicecs would get the same blocking queue, i.e., the 1st one inside OrderedMultiDeviceLoDTensorBlockingQueue.