* add master gradients on static graph

* add unit test for bf16 master grad static graph

* use float16 as v100 test dtype

* only skip GPU which do not support bf16

* use linear layer to test master grad

* 1.push master grad creation before all optimizer ops; 2.remove useless unittest; 3.use a function to create master grad states

* Allow to switch whether to use promote strategy to choose kernel for O2 training.

* Fix comparing error and add unittest.

* support OD level and skip dynamic loss scaling for bf16