[Proposal] Introducing the concept of invalid costs to the IR cost model

It might also make sense to have functions that return costs for vector ops return some sort of LinearPolyBase with ScalarTy being InstructionCost.