Ground plane is always a good idea. Probably not too much difference at these frequencies and impedances, really.
The only place you'd pull back ground plane is around the -in nodes, to reduce parasitic capacitance. An inner copper ground might give you another pF on the node, which is significant for 100MHz+ amps, but won't make a difference here (as you note, the pins are already ~8pF -- though they don't say what that's relative to, be it pin-to-pin, pin-to-ground, or some combination inbetween).
Capacitance on -in isn't necessarily fatal, anyway; I've used it before (or an R+C to ground) to improve RFI hardness, or peaking the frequency response (not so much reducing the phase margin, as optimizing it).
Regarding the VREF node, you can use a series resistor to decouple the op-amp output from the (now able to be much larger) bypass cap. This ruins DC correctness, but if you don't need much or any DC current, that's fine; if you do need it, then you can move the feedback path to the bypass cap as well. Now you need a series R in the feedback path, and a bypass cap from output to -in, to compensate it. The Rout-Cbyp and Rfb-Cfb time constants need to be more-or-less complementary, otherwise you get a big spike in the output impedance, or instability in the amp.
Tim