For very high resolution you will need some kind of chopping or auto-zero for the input stage anyway. A normal high impedance input stage will have way more 1/f noise than the ADC. As these ADCs are really fast, they are more suitable to use just a chopping stage at the input and do the demodulation in software. So no classical chopper OP, but more like the auto zero of DMMs. Measuring zero and the signal one after the other (or changing polarity). This way 1/f noise of the ADC is not that relevant anyway.
Even with chopper OP there will be some extra low frequency noise from thermal effects - not classical 1/f noise, but similar. The other part is the 1/f noise of the reference - at a certain level even a LTZ1000 will add to the noise.
With such a fast ADC, there is not that much analog filtering needed before the ADC and thus no high impedance needed.
Yes, that is one way of doing it. Chopper opamp on the input, analog switches filter after it, and a buffer for the ADC. The buffer is a traditional opamp, the analog switch (2x 1:3, signal, ground, vref ) is switched with a few KHz. After that you need to iron out settling time issues for the buffer, timing and digital magic. I had an analog fronend operating with this principle, the end result was excellent. It was compared to the 3458A at work, they were tracking each other very closely for weeks. I think the result would be a lot better, if it would be a purpose designed data acquisition card. Actually I thought it several time to make my own, probably after my NDA with my ex company expires. The only issue is that the BOM cost is so high, that even a functional prototype is expensive, and I dont have access to a volt nutter multimeter anymore.
I've been thinking about making it a kickstarter, but I actually dont think there are too much volt nutter out there that would buy it. And the ones that do, know (at least parts of it) lot better, so they would rather dispute than buy.