[llvm] [InstCombine] Fix ninf propagation for fcmp+sel -> minmax (PR #136433)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> This makes some sense to me (esp as these are the only possible semantics for phis) but I can't find any wording to that effect in LangRef.

This is news to me. I have always assumed this applied to both inputs like every other operation 

https://github.com/llvm/llvm-project/pull/136433