[llvm] AMDGPU: Avoid using kernels in f16 copysign test (PR #142113)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> Do we have to use entry CC here?

No. The only reason to use amdgpu_ps is to get the return-value-is-SGPR behavior 

https://github.com/llvm/llvm-project/pull/142113