* add gradient merge for DistributedFusedLamb

* use master acc gradient

* fix CI ut

* polish

* remove math_function_impl.h change

* fix test_update_loss_scaling_op.py

* try to fix XPU/NPU CI

* add gm ut