Don't promote fp16 to fp32 (#1796)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/1796

for listed ops in the pass we may not want to use compute type but instead want to use result dtype.

Rationale:

for Add,

`fp16 + fp16` results in `convert to fp32, fp32 + fp32` which may be the desired outcome for either of the ops.

Reviewed By: mcr229

Differential Revision: D53248920

fbshipit-source-id: 0eca43cc8f9a88426cf1904e1e66b4a205d5f5c6