[cherry-pick][hybrid performance] Grad fuse for gradient merge under pipeline mode (#35004) (#35299)

Co-authored-by: Ngongweibao <weibao.gong@gmail.com>

* add retry for HcclGetRootInfo

* refine code

* reduce retry interval

* support class center sample of PartialFC

* copy boost optional.hpp to paddle

* copy boost optional.hpp to paddle

* move directions

* del fluid/utils

* modify .hpp to .h

* move directions

* modify to paddle::optional

* add modification description

* format code stype for the files in paddle/utils

* format code stype

* - disabled caching of layer norm

- fix in compilation

- compilation fix

- transpose caching disabled

- compilation fix

- more compilation fixes

- sum caching disabled

- compilation fix

* - LRN with disabled cache

* lint fixes

* [bug fix] fix unfold negative_size_param

* dygraph support more ctc grad scale

* scale for 1.x

* fix unitest

* fix unitest

* format code

* fix unittest

* fix log info

* unittest cov

* fix format;notest,test=cpu,coverage

* skip ctc_loss egs;test=cpu

* warpctc grad cov;test=coverage

* add dygraph test;test=coverage

* format;test=cpu,coverage

* format;test=cpu

* add api compat;test=cpu

* add cpu test

* rename

* rename

* fix

* fix test

* format

* eigen cpu

* eigen gpu grad pass

* cuda gpu pass

* format

* fix ci

* add inplace passes and tests

* update

* fix use_cuda undefined
fix compile error of op compat

* add more ut

* fix CPU CI error

* check adam unique

* fix mac/windows ci, improve coverage

* fix ci error

* follow weihang's comment

* fix BlockDesc::MoveFrom

* follow qiuliang's comment

* update

* follow huihuang's comments

fix a bug in nlp: text_matching/sentence_transformers when last dim is 1 and reduce mid dim (#34941)

* Fix typos in english docs for diag and diagflat.

* [NPU] Support npu op:(1)arg_min (2)arg_max

* Modify and add unit test cases

* Modify unit test cases

* change bilinear thread for nano and tx2

* change bilinear thread for nano and tx2

* add size npu op

* modify support data type

* no longer use NPU size OP

* remove useless comments, add test case

* fix copyright, remove useless include

* - Added softmax without caching

* - Binary is no longer manually cached

* - Activation onednn caching removed

* - Removed manual caching of activation

* - modified UT

* - fix

* - fix

* - fixes to building

* - fix

* - fix

* - fix to UT

* - Faulty UT workaround

* - approval workaround

* - Fixes after review

* - compilation fixes

* - more lint fixes

* - more fixes after review

* - fixes after another round of review

* - hopefully compilation fix

- compilation fix

* Add NPU kernel for nearest_interp op

* Add grad op

* Modify codes according to the review comments

* Modify codes according to the review comments

* add unique_consecutive_op

* add unique_consecutive_op

* add unique_consecutive_op

* add unique_consecutive_op

* add unique_consecutive_op

* add unique_consecutive_op

* add unique_consecutive_op

* add unique_consecutive_op

* remove unity build

* add unique_consecutive op

* add unique_consecutive op

* add enable static

* add noqa

* add space line

* add default case.

* add comma

* add space line

* modify unique_consecutive unittest

* optimize ut coverage

* rebase develop

* improve coverage

* update en docs

* update en docs

* update en docs

* update en docs

* update en docs

* update en doc

* support margin loss (arcface, cosface, sphereface)

* Support NPU OP hard_swish and hard_swish_grad

* Support NPU OP hard_swish and hard_swish_grad

* add the unittest to compare the result between npu ans cpu

* format the prompt of exception

* replace Min and Max op by ClipByValue op

* fix the precision problem for fp16

* Using HardtanhGrad to improve performace

* add p_norm_op_npu

* remove p_norm_grad op

* update

This reverts commit 0a5c99e8.

This PR adds fused transformer related files defining c interface including class, function etc..

* Fix safety-bug of functional.linear

* Fix safety-bug of functional.linear

* Fix safety-bug of functional.linear

* Fix safety-bug of functional.linear

* [NPU] Support npu op expand_v2 and expand_v2_grad

* [NPU] Support npu op expand_v2 and expand_v2_grad

* [NPU] Support npu op expand_v2 and expand_v2_grad

* update test_expand_v2_op_npu.py

* update test_expand_v2_op_npu.py

* modify expand_v2_op_npu.cc

* modify expand_v2_op_npu.cc

* - Added softmax without caching

* - Binary is no longer manually cached

* - Activation onednn caching removed

* - Removed manual caching of activation

* - modified UT

* - fix

* - fix

* - fixes to building

* - fix

* - fix

* - fix to UT

* - Faulty UT workaround

* - approval workaround

* - Fixes after review

* - compilation fixes

* - more lint fixes

* - more fixes after review

* - fixes after another round of review

* add set_value_grad op

* add unittest.

* polish unittest.

* polish code.

* support cuda kernel

* polish code according to CI

* polish code.

* polish code

* remove *.pyc

* polish code.

* add unittest to improve coverage.

* polish code.