* support elementwise add, activation, matmul on Baidu Kunlun
* test=kunlun

* minor
* test=kunlun

* reconstuct the xpu directory
* test=kunlun

* minor
* test=kunlun

* minor
* test=kunlun

* minor
* test=kunlun

* minor
* test=kunlun

* minor
* test=kunlun

register log double grad kernel for cpu and cuda

* fix test_dist_fleet_heter_ctr & peformance update

* fix pool bug, test=develop

* fix coverage,test=develop

* fix bug, test=develop

* refine vision models

* add static check on data()

* follow comments

* fix ut

* add xpu in heter mode
test=develop

* BOOST_CONST_GET; PADDLE_THROW
test=develop

* code style
test=develop

* code style
test=develop

* code style
test=develop

* refine
test=develop

* refine
test=develop

* refine
test=develop

* refine code
test=develop

* fix bn double grad, test=develop

* update, test=develop

* try to fix paddle save load test

* open paddle save load

* replace dataloader

* remove dataloader

* add fp16 for matmul

add abs support double grad for the api 2.0

* [buf fix]:fix unittest test_activation_op error

* split long-time unittests to smaller ones

* rename some unittests

* fix unittest_group_norm_op_v2, test=develop

* fix unittest_group_norm_op_v2, test=develop

* Add pass compatibility check for four mkldnn pass, test=develop

* make places of DataLoader optional. test=develop

* test=develop, bug fix

As the title, decrease random failure probability for test_parallel_executor_mnist

The old code set larger delta when comparing reduce and all reduce, but didn't set all. I added it.

On my linux machine, I run 100 times, no failure occurs. In addition, we only saw this random failure on CI two times since I worked. I thought it was rare and I just increased the delta.

* Remove dependency on nltk for paddle __init__.
test=develop

* Remove nltk.movie_reivew sentiment dataset to remove dependency on nltk.
test=develop

* reproduce summary api