fixes to layernorm emulation (#40422)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/40422
fix the remaining differences to the emulation of fp16 layernorm
Test Plan: unit test of layernorm
Reviewed By: venkatacrc
Differential Revision: D22182849
fbshipit-source-id: 8a45c21418517d65d7a41663d5ad2110d6b4677a
