Scale decoding architectures to lower parameter counts and to fit on smaller GPUs

[reesekneeland]

MindEye 1 and 2 in their default training/inference configurations require an A100 to use. There has been other recent work exploring a reduction in parameter counts that could be valuable to implement, in service of our tertiary goal of making these decoding algorithms more scalable and easier to use. This is also a good item for people with limited compute (no A100s) to work on.

Lite-Mind paper: https://arxiv.org/html/2312.03781v1

Other easy things:

• Don't load all of the images onto the CPU
• Smaller batch sizes
• Disable unnecessary modules (captioning module, etc)