DNL is not a measure of the statistical deviation of the difference of two voltages. DNL i s the deviation of the step from one voltage to the nex, compared with the average step of the DAC. Then, on top of that you can ask about the statistical variation of the DNL or the error in the step.
In that respect, for thermometer DACs, you always step from one code to the next by adding one unit element - current in your example. And you compare it with the average value of that unit element, which the average LSB step. DNL=(I-Iavg)/Iavg. To find the variance you simply have
var[(I-Iavg)/Iavg] = var(I/Iavg)=var(I)/(Iavg^2) and thus sigma=sigma(I)/Iavg