[ValueTracking] call nnan/ninf noundef cannot return poison (#239912)
While working on #239458, I noticed that ValueTracking had inconsistent
semantics for call instructions with both nnan/ninf and noundef.
canCreateUndefOrPoison would return true, but
isGuaranteedNotToBeUndefOrPoison would also return true.
This effectively decides on one of the two semantics, saying that `call
{nnan,ninf} noundef` cannot return poison (effectively nnan/ninf getting
evaluted before noundef). This keeps `nnan noundef` consistent with
`nofpclass(nan) noundef` and makes `noundef` actually useful with FMF.