Estimating the bias of CX gates via character randomized benchmarking