You are talking about two completely different things, the only link between them is the bias voltage generates an artificial d.c. offset to the input to the ADC, which the filter is intended to remove so as to leave what the ADC would have measured if it could accept negative-going inputs.
The buffered bias is intended to supply the bias voltage to several inputs from one source, thus potentially saving components whilst providing a more stable voltage.
The problem with a filter - any filter - is striking the balance between the time it takes to settle after switch-on, and the unwanted removal of the lowest frequency in the desired band - in our case in Europe, the 50 Hz mains frequency. If you go past first order filters, then you need to take processing time into account too.-
There are at present two techniques in use, a third was used initially.
The first to be used was a high pass filter. This removed the d.c. component, but required a comparatively long settling time because it could not be initialised to somewhere near the expected value.
That was replaced in emonLib by the present technique, where the d.c. component is extracted by a low-pass filter, then subtracted from the signal to leave only the desired a.c. with no offset.
In emonLibCM, the nominal offset is removed first, then the average signal over the whole sampling interval computed and the remaining d.c. component is extracted after calculation of the average (for the power) or the rms average (for the voltage and current).
But, if you use the buffered bias and you have a spare ADC input, then you can measure the bias voltage directly with the spare input and simply subtract that from the signal - what you have then is the emonLib method without needing to use a software filter.