[CostModel][x86] remove cost-kind predicate for intrinsic costs

We model cost as number of instructions / uops, so it does not
make sense to treat size/blended costs any differently than
throughput.