[clang] [AMDGPU] Use generic builtins for `wave_reduce` ops (PR #179589)

[Shilei Tian via cfe-commits]

shiltian wrote:

> Why? Again, the algorithm itself is generic, you're just making life slightly harder for the user for no clear benefit, why would a HLL user care that the **intrinsic** that gets selected is reduce_fmin or reduce_fmax?

I guess it is because `fmax` and `fmaximum` are two things for floating-point values, and calling them `max` and `min` is not definitive enough.

https://github.com/llvm/llvm-project/pull/179589