[llvm] [AMDGPU] Support double type in atomic optimizer. (PR #84307)

[Jay Foad via llvm-commits]

jayfoad wrote:

It looks like this patch only affects "noret" cases with uniform input values, so it does not compute any reductions or scans, so it does not exercise either the Iterative or the DPP algorithm.

https://github.com/llvm/llvm-project/pull/84307