[unified arch] Cache the outputs of the vision encoder

[neilmehta24]

Start tracking the tokens and cache in cache_wrapper. When we receive a followup prompt, we now no longer reprocess the images.
Cases:

• The hash of the input images changes --> Always do full reprocessing
• The prompt is extended --> Do text-only prompt processing
• The prompt is trimmed --> Trim the cache and do text-only prompt processing
• The cache cannot be trimmed --> Full vision reprocessing

I added two tests, one for SWA caches which often cannot be trimmed, and one for non-SWA caches, which are usually always trimmable.

Note that there are still opportunities for improvement. Namely, we could be caching the embeddings per image so that we can selectively re-use the embeddings. This can be added as a feature in a future PR, in cases where the cache cannot be trimmed.

Note that this doesn't cache images going through the non-unified stack, that can be added in a future PR.