For any FET, there should be one "magic" point where
Vgs is relatively constant. This beats the subthreshold slope
change against the VT change for a net null. I would not expect
it to be fabulously consistent, but play around with a FET at a few
different current densities, across temp and you may see
what I'm talking about.
Then you only have the small problem of getting that accurate,
constant-vs-temp current instead. Heh.
Another option is to make the bias current decrease at a proper
rate w/ temp. If you had a bandgap reference you could do
this closed loop, perhaps.