* Add casting initializers for bf16 training

* Changes after review

* Correct test and add comment

* support xpu inference with analysis predictor, test=develop

* merge the cmake of the xpu toolchain, test=develop

* add c-apis, test=develop

* fix a bug in extern_xpu, test=develop

* default open dygraph mode

* fix CI-Mac

* fix Mac-CI other unittest file

* fix CI-Py3

* fix test_communicator_geo and test_buffer_shared_memory_reuse_pass

* add enable_static to fix CI-Py3

* add enable_static to fix CI-coverage

* delete try except

fix test_word2vec float's equality bug

add python inference api

test=develop

test=develop

This can cause a lot of problem:
1. Wrong operator implementation, Op can get a wrong type without failure.
2. Anytype can be Get without defined in VarType.

Also fix wrong STEP_SCOPE usage.

test=develop

* add lod_tensor utility python module

* add lod_tensor test code

* add more lod tensor tests

* modify word2vec example code using new api

* add comment

In this way we can build and test using WITH_FLUID_ONLY flag being set
to ON.

- move paddle.v2.dataset,reader to paddle.dataset,reader
- remove unused code (which depends on v2) in paddle.dataset,reader

* merge fluid dist tests

* update cmake

* Generated functions can take `*args` as inputs.

Use a new scope for inference in python unittest to avoid changing the value of variables for training.

* Add unit-test and example

* Fix type error

* Fix unit test cases

* Fix init error for cudaplace

* Change unit-test options

* Fix parallel.do with batch norm

* Change log level

* CopyShare AllPlaces

* disable nccl test

* bring back parallel_do test

* Polish sum_op support SelectedRows in_place