Add a test to load a full model and allow to to post-train from C++

* add multi_label UT

* rename, fix typo, add fuse_statis check

Enable MKL-DNN in Analysis Predictor

little update to conv relu fuse pass (fix)

disable nccl2 test on CPU-only mode

Revert "Revert "Add Python Callstacks when Op::Run error (#12759)""

Also fix MKL-DNN pooling integration for ceil mode

* mac ci make install fix

* fix build

* update

* update

[Urgent] Add out params to some apis

clean a few more kwargs

Hide parameter

only mode

Move trainer to contrib

fix table grad send sync_mode

* generic cmake

* Update generic.cmake

* add nccl2 dist api

* update apispec

* update

* update api spec

* add guide

* "fix doc"

* Update windows_inference.md

Looks like there is a little problem in markdown format writing of head lines

Remove kwargs in softshrink

fix links in readme