[llvm] Avoid maxnum(sNaN, x) optimizations / folds (PR #170181)

[James Y Knight via llvm-commits]

jyknight wrote:

I sympathize with why this was done, but I really don't the precedent of disabling pre-existing constant-folding as an answer to an open debate as to LLVM IR semantics. I don't want to push for undoing it now, but I hope this will be reverted once the discussion thread completes.

https://github.com/llvm/llvm-project/pull/170181