vectorize rounding ops (#41439)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/41439
use RoundToFloat16 on arrays
Test Plan: layernorm unittest
Reviewed By: venkatacrc
Differential Revision: D22540118
fbshipit-source-id: dc84fd22b5dc6a3bd15ad4ec1eecb9db13d64e97