Behavioral macromodeling of high-speed drivers via compressed tensor representations