* add paddle.async_save to reduce time cost by checkpoint saving

* adapt save_for_auto_inference to paddle.async_save

* modify UT

* modify UT

* fix on cpu only version

* revert commit on save_auto_inference

* fix threading