[llvm] [AMDGPU] Add getCmpSelInstrCost override to re-enable SimplifyCFG speculation for vector types (PR #208043)

[Frederik Harwath via llvm-commits]

frederik-h wrote:

@michaelselehov As you suggested to me elsewhere, the cost kind check was too broad. I restricted it to `TCK_SizeAndLatency` and I added a test that documents the (non-)effect on the unroller .

https://github.com/llvm/llvm-project/pull/208043