New model gev outperforms jev and takes images as input too

I've made a new model `gev` that outperforms `jev` in accuracy on JevBench and it takes image input too (unlike jev), and gev can be faster too!

also this AnyEval we made is great. can set up any model vs model comparisons for any benchmark

full traces. cryptographically signed

New model gev outperforms jev and takes images as input too (anyeval.com)