[llvm] [AMDGPU][NFC] Templatise and roundtrip gfx12_asm_vop3_dpp16.s (PR #203953)

[Shilei Tian via llvm-commits]

shiltian wrote:

> > it took 150-180 seconds, and with the patch it takes 260-300 seconds.
> 
> Since the test now effectively combines the work of what previously were two separate tests (which is a significant improvement in convenience to us), the change in the numbers make perfect sense.

Does it make sense to split them to improve testing parallelism?

https://github.com/llvm/llvm-project/pull/203953