fix bug in dist train on hs, test=develop

Add cudnn 5 support

support customized kernel selection

Optimize hs op

test=develop

test=develop

fix a const_cast and avoid using stale program.

* test dist ut fixes test=develop

* fix cmake

* for test

fix mac ci test step, test=develop

* expose square
test=develop

* fix activation
test=develop

* Add square API
test=develop

* add necessary op

* code refine

* fix API.spec
test=develop

* fix unit test
test=develop

* add unit test sparse_grad_clip
test=develop

* fix API.spec
test=develop

* remove mac test for test_gradient_clip
test=develop

* remove selectedrows_mul_tensor
test=develop

My cool stuff

* wip multi process multi gpu dist training

* workable for p2p

* update test=develop

* change back env name test=develop

* fix alloc init

* fix cpu build test=devlop

* fix mac tests test=develop

* refine code

* refine test=develop

enable API check for readers

* layer fixes test=develop

* follow update test=develop

, test=develop

, test=develop

add more files to protected file list

* fix api format and examples

test=develop

* Update executor.py

test=develop

* Update nn.py

* Update nn.py

test=develop

* Update nn.py

test=develop

Fix sequence_mask_op InferShape

Fix the bug that profiler cannot trace the nccl allreduce operator

test=develop

Revert "Add EstiminateFlops"

Add YOLOv3 loss operator for YOLOv3 model

add mac ci check on import

test=develop

test=develop