Avoid type promotion pass for fp16 weights (#1694)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/1694

We are inserting `_to_copy` ops to `mul` where we promote the dtypes of input tensors to fp32. This behavior is unwanted for fp16 models. Here I'm adding an option to `to_edge()` to disable this type promotion pass.

Reviewed By: angelayi

Differential Revision: D53033258

fbshipit-source-id: cf1b61f0a687ac07e4f937f40c3db1be29f26dcf