Created by: wilhelmzh

1. make reset auc states in infer more elegant
2. add scope_guard in ctr/dnn infer.py
3. use fleet instead of directly using transpiler.
4. add load model before local distributed training.