Created by: jacquesqiao

Merged PR:

1. Optimize sum op https://github.com/PaddlePaddle/Paddle/pull/13818
2. prefetch optimize https://github.com/PaddlePaddle/Paddle/pull/13887
3. Fix lookup table at CPU Reduce strategy https://github.com/PaddlePaddle/Paddle/pull/13985
4. update shape from int to int64 https://github.com/PaddlePaddle/Paddle/pull/14065 to support embedding wide larger then 2000000000.
5. add flags for control the thead num for pserver https://github.com/PaddlePaddle/Paddle/pull/14060
6. use fake init op to init table param on trainer side https://github.com/PaddlePaddle/Paddle/pull/14096
7. fix unittest for CPU Release 1.1 https://github.com/PaddlePaddle/Paddle/pull/14098

Need discuss

1. ctr c++ reader https://github.com/PaddlePaddle/Paddle/pull/13983