[llvm] [AMDGPU] Model the cost of the expanded integer to/from floating point casts (PR #223572)

[Dmitry Sidorov via llvm-commits]

MrSidims wrote:

>  but it's probably better to fix the scalar counterpart. Maybe as a follow-up.

Preparing few fixes, partially in this PR, partially in a new PR. Thanks!

https://github.com/llvm/llvm-project/pull/223572