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

[Matt Arsenault via cfe-commits]

arsenm wrote:

> So, basically, if we wanted to upstream all of this, it'd need to live somewhere and I'd assume `libclc` would absorb it.

I think the actual algorithmic stuff should go into libm. Ideally we would reduce libclc to the OpenCL glue 

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