[DAGCombiner] Relax nsz constraint with fp->int->fp optimizations (#164503)
`NoSignedZerosFPMath` isn't a hard requirements and in some contexts we
can still apply the truncation without worrying. For example, in cases
where the users of this sequence are overwriting the sign-bit (fabs) or
simply ignoring it (fcmp).
I think the same logic can be applied elsewhere for other DAG
optimizations.