Another interesting approach, since you are always measuring something with a large, fixed polarity (2.5V - 3.65V) would be to "subtract out" a reference voltage, allowing you to gain up the remaining difference and get higher signal before you saturate your ADC. Basically it would allow you to use your ADC's input range more fully by "zooming in" on the voltage range of interest, not wasting range on the first ~2.5V which is always present in LiFePO4 batteries.
There are quite a few non conventional approaches that offer some intriguing possibilities here.
My preferred method is to use a voltage to frequency converter for measuring cell voltages.
Then just make a completely digital frequency measurement.
If you require a completely floating voltmeter for cell voltage measurement, data transfer requires only one opto isolator for the frequency.
I have thirty cells and over a hundred volts across the battery stack, so any kind of off the shelf multiplexer chip is not an option. But that may not be a problem you have... But this voltage to frequency system offers a lot of other advantages as well.
If the VCO frequency is kept fairly high, and the frequency measurement sampling period suitably long, you can get very high resolution, as many bits as you want.
Any noise on the cell voltage created by inverter ripple current, just creates frequency modulation. That all averages out over the frequency measurement period.
By allowing the frequency counter to overflow, or just ignoring the most significant bit or bits, you can subtract out a major portion of the measurement scale such as half or three quarters. Or if you play with the numbers, any other proportion as suggested in the above quotation^.
If the frequency measurement period is very closely tied to the 100/120Hz inverter ripple frequency, so it closely matches a whole integer number of half inverter cycles, inverter noise rejection will be even better. It does not need to be phase or frequency locked, just fairly close, easily achievable with a quartz crystal reference.
Absolute accuracy is not really required, if 3.45 volts is actually out by several millivolts its not a big deal. But cell to cell voltage differences need to have suitably high resolution, and be very repeatable and noise free.
Slow temperature drift is not going to be a problem, but wide band system noise certainly will be. Slowing down measurements with restricted bandwidth does not work.
If it takes 30 seconds between readings to reach 1mV final settling accuracy, its all going to be far too slow to measure really small cell voltage differences with a large number of cells. If it all responds fairly fast, its going to be noisy where there may be several different high power solar controller pwm systems hanging off the battery as well as the inverter.
You need to be a bit sneaky how you get around the noise problem. The voltage to frequency method has turned out vastly better than any other system so far. I have tried quite a few different approaches to this over the years.
Its possible to very easily get 1mV or 2mV per bit completely noise free without any subsequent averaging in software, scanning at one cell per second rate.