* bugfix (#45332)

* dist embedding support lookup table v1

* add unitest

* customize wait_comm

* group gradients

* bugfix

* update program

* adapt lazy init and fix pass

* add unittest

* update comment

* fix amp and sharding

* remove clip_by_norm

* bugfix (#45332)

* customize wait_comm

* support high order differential with data parallel overlap

* update unitest

* bugfix

* remove scaling

* support rescale_grad opt

* add unitest

* bugfix

* remove scaling

* support rescale_grad opt