* Some improvements to support bert mixed precision training

test=develop

* Revert the cast in layer_norm

test=develop

* Simplify the compare op for CPU.

* Use asynchronous tensor copy in reshape_op's kernel.

* Optimize while_op for test, avoiding creating variables every time.
test=develop

* Enable the cache of kernel type and kernel function.
test=develop

* Enable profiling with gperftools.

* Remove flags for testing, and fix the linking error.
test=develop

* Delete the codes of ChooseKernel.
test=develop

* Fix bug when preparing ExecutorPrepareContext for while_op.

* Fix missing depending on grpc libraries.

* Remove the redundant print.
test=develop

* Follow comments.

* Remove the codes related to prepare the ExecutorPrepareContext for while_op.
test=develop

test=develop

test=develop

* Add intermediate

* fix flatten/squeeze/unsqueeze

* Considering compatibility issues, we could not fix the origin op

* follow comment

* reset the shape of XShape

* "remove inplace in single op"

* "fix ci"

* "add transpiler case"

* fix conflict

* "fix reshape"

* "delete reshape inplace attr"

* "follo the comments"

* "rerun ci"

* Shrink reshape_op library size
* User can register a standard C++ functor as a op kernel

Do not use ctor

* Reduce line of codes.
* We can use virtual function for Maker now.
* The implementation does not care what maker holds, it is easier to
refactor later.

* script to add semicolon

* fix typo

* Move framework.proto to proto namespace

* Fix compile

* Fix compile

* Fix Compile

* polish r operator docs

* fix on naming convention

Use posix_memalign to allocate aligned memory, since some SIMD instructions require the alignment of memory accesses.

* add batch_norm_layer

* add img_conv_group layer and test

* add check to Tensor.type()

* forward can run

* with backward

* change label data time from int32 to int64

* refine code

* follow comment