[libcxx] [llvm] [libc++][atomic_ref] Use __atomic_fetch_{add,sub} builtins on floating-points whenever possible (PR #135685)

[Damien L-G via llvm-commits]

dalg24 wrote:

The CI failures are unrelated as far as I can tell.

The `generic-{msan,tsan}` builds fail when configuring Google benchmark https://github.com/llvm/llvm-project/blob/32180cf9f9eb921a793bb208cf3bbfb1b86ee2ae/third-party/benchmark/CMakeLists.txt#L314-L316

The `generic-no-localization` build fails some "libcxx mangled names" test which does not seem to use anything that changed in this patch
https://github.com/llvm/llvm-project/actions/runs/15928686318/job/44937568088?pr=135685

https://github.com/llvm/llvm-project/pull/135685