Created by: dzhwinter

fix https://github.com/PaddlePaddle/Paddle/issues/9099 every minibatch sequence_pool and sequence_pool_grad operator have a ~8x time acceleration. for example, the sequence_pool op enhanced from 0.815583 -> 0.119373 the sequence_pool_grad enhanced from 0.579614 -> 0.0830757

before optimize

Event                                     Calls       Total       Min.        Max.        Ave.
thread0::sum                              72772       6579.74     0.013088    3.43046     0.0904158
thread0::mul_grad                         25928       4135.29     0.049952    4.4888      0.159491
thread0::sequence_softmax_grad            2344        3067.88     0.05872     95.0493     1.30882
thread0::sequence_softmax                 2344        2617.72     0.04976     17.122      1.11677
thread0::mul                              25928       2260.75     0.038624    8.36944     0.0871933
thread0::sequence_pool                    2380        1941.09     0.045984    89.9217     0.815583
thread0::sequence_expand_grad             2344        1730.34     0.05296     8.10054     0.738201
thread0::sequence_pool_grad               2380        1379.48     0.03824     137.793     0.579614

after optimize

thread0::sigmoid_grad                     7035        304.461     0.024032    89.5161     0.0432781
thread0::sequence_pool                    2381        284.226     0.053984    56.5243     0.119373
thread0::sigmoid                          7035        214.732     0.02448     3.57146     0.0305233
thread0::tanh                             7071        214.441     0.023712    1.59734     0.0303269
thread0::tanh_grad                        7071        206.762     0.023328    0.1432      0.0292408
thread0::sequence_pool_grad               2381        197.803     0.057408    0.934464    0.0830757
thread0::adam                             936         187.986     0.024384    1.28653     0.20084