storing datasets for computer vision evaluation

[rbavery]

currently text eval datasets are stored as python strings

for computer vision, we will need to provide examples differently. One idea: we just spoof arrays in memory to resent 3 channel images. 4 channel images.

[rbavery]

we can start on this once we have an exported torchscript model #21