Enhance hierarchical sigmoid op

* optimize bilinear_tensor_product
* add set zero to set grad to 0.

EltwiseMul: Changes from previous PR

Fix the invalid list operation on save_inference_model function.

Add ctr reader

fix the build issue on manylinux1

test=develop

test=develop

test=develop

test=develop

Protect important header files.

Revert "fix transfer cache thread_local bug (#14581)"

test=develop

Windows/online

* Fix truncated normal.

* Fix.

* Make nce support more distribution.

* Fix API.spec.

* Fix python API.

* Fix.
test=develop

* Fix API.spec
test=develop

* Fix sampler.

* Fix order of arguments in python API.
test=develop

* NCE add selectedrows support

* NCE update weighted sampling

* fix bugs in nce_op, and assign_value_op optimized

* fix bugs in nce_op, revert assign_value_op

* nce_op optimize

* nce_op optimize

* nce_op optimize

* add selectedRows test later

test=develop

* add selectedRows supported

* add selectedRows supported

test=develop

* add selectedRows supported

* add nce selectedRows supported, test=develop

* add nce selectedRows supported

* add nce selectedRows supported, test=develop

* fix height in nce, test=develop

* add ut

* add ut, test=develop

* make AutoGrownIndex inline
test=develop

* fix tinny error, test=develop

Fix Mac unittest bug

add some activation api examples.

test=develop

test=develop

clean inference include files

test=develop