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

[Ralf Jung via llvm-commits]

RalfJung wrote:

I hope this will only be reverted once frontends have a viable alternative to switch to if they want the semantics that used to be implemented by `minnum`. At least the last thing I heard is that `minimumnum + nsz` is not viable at the moment; I am still trying to gather up-to-date information on that.

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