Add multi-subject pretraining to MindEye1

[reesekneeland]

MindEye1 was one of the best performing models on imagery in the initial benchmarking. We should test adding the shared subject latent space and multi-subject pertaining step to see if we can get comparable performance while reducing the necessary amount of training data.

[reesekneeland]

With the implementation of versatile diffusion into the ME2 architecture, this will not be necessary