[llvm] [ValueTracking] Fix computeKnownFPClass handling of nsz (PR #186315)

[James Y Knight via llvm-commits]

jyknight wrote:

As I think what the Archer report's errors have also indicated: this patch seems to expand zero classes in the wrong place.

It ought to be modifying the known values on _inputs_ to an nsz operation to encompass both positive and negative zeros when either are present, rather than to the _outputs_ of the nsz operations.

https://github.com/llvm/llvm-project/pull/186315