NPNs (at least in an "analog" process where they are a
valued element and processing cares about mismatch)
will tend to have a better natural matching in a smaller
area than a NMOSFET.
You can do a chopped and filtered bandgap with good
results, provided that you can tolerate a high output
impedance and some chop-frequency residue. In a
clocked circuit this is probably true. In a low noise
continuous-time amplifier based circuit, probably not.
Back when I did bipolar products we'd get 1mV-range
natural offsets out of NPN front end op amps, few
mV out of ones that used a PNP follower in front of
the gain stage. You need a lot of MOSFET area to
match that performance (in today's terms) - although
back in the day, it was > 10x10um emitters and maybe
it's not that far different after all.
Your foundry should provide you matching data for
both types of device, and just going by the front
end current density of interest and the geometry
it would take for the matching you want, you could
make a value judgment.