fix reduce op in nlp: text_matching/sentence_transformers when last dim is 1 and reduce mid dim

[AnnaTrainingG]

PR types

Bug fixes

PR changes

OPs

Describe

fix a bug in nlp: text_matching/sentence_transformers when last dim is 1 and reduce mid dim

bug现象：

global step 10, epoch: 1, batch: 10, loss: nan, accu: 0.42500, speed: 12.92 step/s
global step 20, epoch: 1, batch: 20, loss: nan, accu: 0.42188, speed: 23.03 step/s
global step 30, epoch: 1, batch: 30, loss: nan, accu: 0.42292, speed: 23.87 step/s
global step 40, epoch: 1, batch: 40, loss: nan, accu: 0.43906, speed: 22.53 step/s
global step 50, epoch: 1, batch: 50, loss: nan, accu: 0.43562, speed: 22.56 step/s
global step 60, epoch: 1, batch: 60, loss: nan, accu: 0.43281, speed: 23.17 step/s
global step 70, epoch: 1, batch: 70, loss: nan, accu: 0.43661, speed: 21.93 step/s
global step 80, epoch: 1, batch: 80, loss: nan, accu: 0.43789, speed: 23.65 step/s
global step 90, epoch: 1, batch: 90, loss: nan, accu: 0.43368, speed: 23.62 step/s
global step 100, epoch: 1, batch: 100, loss: nan, accu: 0.43188, speed: 22.27 step/s

正常运行结果：
global step 10, epoch: 1, batch: 10, loss: 0.77060, accu: 0.56875, speed: 12.83 step/s
global step 20, epoch: 1, batch: 20, loss: 0.64212, accu: 0.58437, speed: 23.04 step/s
global step 30, epoch: 1, batch: 30, loss: 0.63177, accu: 0.60729, speed: 23.92 step/s
global step 40, epoch: 1, batch: 40, loss: 0.61330, accu: 0.60703, speed: 22.58 step/s
global step 50, epoch: 1, batch: 50, loss: 0.63015, accu: 0.62125, speed: 22.58 step/s
global step 60, epoch: 1, batch: 60, loss: 0.65486, accu: 0.63281, speed: 23.30 step/s
global step 70, epoch: 1, batch: 70, loss: 0.60717, accu: 0.63705, speed: 21.97 step/s
global step 80, epoch: 1, batch: 80, loss: 0.61400, accu: 0.63984, speed: 23.68 step/s
global step 90, epoch: 1, batch: 90, loss: 0.48885, accu: 0.64306, speed: 23.56 step/s
global step 100, epoch: 1, batch: 100, loss: 0.58543, accu: 0.64281, speed: 22.10 step/s

修复后：
global step 10, epoch: 1, batch: 10, loss: 0.77057, accu: 0.56875, speed: 12.79 step/s
global step 20, epoch: 1, batch: 20, loss: 0.64191, accu: 0.58437, speed: 22.83 step/s
global step 30, epoch: 1, batch: 30, loss: 0.63172, accu: 0.60729, speed: 23.88 step/s
global step 40, epoch: 1, batch: 40, loss: 0.61274, accu: 0.60703, speed: 22.50 step/s
global step 50, epoch: 1, batch: 50, loss: 0.63057, accu: 0.62062, speed: 22.54 step/s
global step 60, epoch: 1, batch: 60, loss: 0.65494, accu: 0.63229, speed: 23.25 step/s
global step 70, epoch: 1, batch: 70, loss: 0.60811, accu: 0.63661, speed: 21.94 step/s
global step 80, epoch: 1, batch: 80, loss: 0.61390, accu: 0.63984, speed: 23.60 step/s
global step 90, epoch: 1, batch: 90, loss: 0.48840, accu: 0.64306, speed: 23.63 step/s
global step 100, epoch: 1, batch: 100, loss: 0.58263, accu: 0.64250, speed: 22.28 step/s

现象说明： 模型reduce中间维度的时候如case：【32 33 1】 reduce_dim = 1 模型运行结果为nan
原因分析： 在进行代码整合的时候将kReduceALL 类型删除，并且增加了reduce.left_num == 1 调用cubReduce的判断，
但是在进行ReduceConfig中进行block、grid设置时，更新了reduce.left_num，导致reduce_left_num等于最低维1， 进入cu bReduce调用，使得最终结果计算错误。

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[xingfeng01]

LGTM

[ZzSean]

LGTM