* support elementwise_pow bfloat16

* add only_check_prim parameters in check_grad

* modify unit test

* fix floor test

* fix sigmoid bfloat16 test

* norm some ops prim test

* add unit16 for sqrt