Created by: bingyanghuang

Reported from Baidu that when tested with the large dataset, fused_embedding_seq_pool will cause the NAN error. @Sand3r- helped review the JIT code and he has some questions on this issue and comments in this code. Questions regarding this issue:

1. Does the related JIT code regard forward or backward passes?
2. On what CPU does this code produce crash?
3. Is the code responsible for issue located here? https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/jit/gen/embseqpool.cc
4. How to reproduce the problem? On what model/unit test can one experience this issue? If the model where the NANs occur is secret, we should at least get Input buffers along with their dimensions and attribute settings of the Op.