[libclc] [libclc] Use unchecked division for f64 AMDGPU reciprocal (PR #203809)

[Wenju He via cfe-commits]

wenju-he wrote:

@jhuber6 We have the same problem in our downstream code (will be upstreamed), like group sort, cluster variants and matrix related. I asked the similar question to @frasercrmck before. His answer was
`
I've thought about the problem but I haven't quite worked it out. If libclc and libspirv had a better "story" in upstream clang it would be easier. Right now it's purely an OpenCL library, so no. But with SPIR-V builtins, potentially the project could be thought of as a library for builtins of languages in the OpenCL "family" - SYCL and SPIR-V included. If we can argue for that, then I don't see why other builtins can't belong there.
`
In general, I'm open to putting them in libclc and evolving it as a library of builtin functions written in OpenCL C.

https://github.com/llvm/llvm-project/pull/203809