* support bf16 in AmpOptimWrapper

* add docstring

* modify docs

* add unittests for bf16 in AmpOptimWrapper

* fix type

* fix to pass ci

* fix ut skip logic to pass ci

* fix as comment

* add type hints

* fix docstring and add warning information

* remove check for pytorch>=1.6 in unittest

* modify unittest

* modify unittest

* remove torch.float32 && torch.float64 from valid dtypes

* fix as comments

* minor refine docstring

* fix unittest parameterized to pass CI

* fix unittest && add back torch.float32, torch.float64