Benchmark current MindEye2 architecture on NSD-Imagery using varied training data

[reesekneeland]

Only the final 40 hours version of MindEye2 has been benchmarked on NSD-Imagery. We should evaluate the models trained on the following numbers of hours of training data: [1/2, 1, 2, 3, 5, 10, 20, 37, 40]. This will likely need to wait until we have NSD-Imagery on the stability cluster.

[reesekneeland]

Model inference for the versatile diffusion model has been completed by @ckadirt, needs quantitative benchmarking. We should also run the original ME2 models here