[clang] [llvm] [Clang][NVVM] Support `-f[no-]cuda-prec-sqrt` and propagate precision flag to `NVVMReflect` (PR #134244)

[via llvm-commits]

Lai-YT wrote:

> At first glance, it seems like a good idea to allow people to choose how their square roots are lowered on a per-function level: some code cares about precise square roots, some doesn't, and you should be able to make choices on a case-by-case basis.
> 
> But looking at the code structure, maybe cuda doesn't allow that?

Yeah, AFAIK, CUDA—or more precisely, NVCC—only allows this to be specified at the module level. 😣

https://github.com/llvm/llvm-project/pull/134244