Batch processing should support caching post-foundation model representations

[isaacrob]

For example, if we're using the classifier on a dataset, we should have an option to cache all the CLIP embeddings so that we can run a new classifier config against the dataset very quickly.