[llvm] [polly] [InstSimplify] Fix incorrect poison propagation when folding phi (PR #96631)

[Yingwei Zheng via llvm-commits]

dtcxzyw wrote:

> @dtcxzyw Could you check this on llvm-opt-benchmark first? I'm wondering whether this introduces regressions that can be recovered based on more context.
> 
> The motivation for doing this _now_ after ignoring this for so long is that we're definitely going to lose this when we stop using undef for uninit loads, so I'd like to know what the impact of just this part is going to be.

Done.

https://github.com/llvm/llvm-project/pull/96631