[Speed] sequence_pool op need to be enhanced

[dzhwinter]

Every single call to sequence like op, it will produce a bunch of kernel calls. It should be enhanced.

thread0::array_to_lod_tensor                2           62.4829     30.9197     31.5633     31.2415     3262.35     0           327.645
thread0::mul                                763         61.21       0.038688    0.71024     0.0802228   642.303     0.000488281 25.7815
thread0::sequence_softmax_grad              69          57.409      0.04448     2.6696      0.832014    4320.91     0.000488281 0.0126953
thread0::lod_tensor_to_array                2           56.6311     27.8057     28.8253     28.3155     810.032     5.60864     327.656