Deploy and improve the streamlit demo app

[metazool]

There's a streamlit UI in here (borrowed from the one for language model embeddings search

• Uses a collection of image embeddings from this fork of the scivision plankton model held in chromadb
• Interface to pick a plankton image at random and return the closest, by cosine similarity
• Interface to do k-means clustering of the embeddings, with tweakable parameters, and visually inspect the results

Originally written for #8 - a demo of using image embeddings to explore features across an image collection without more labelling or training. Hopefully useful as a method for picking out the detritus cluster for a "good enough", cheap to make binary classifier - see #32

[metazool]

Started a hello world project over here to test out the continuous deployment from git feature of Posit Connect - I know it's a sidetrack but it would be a useful pattern to have!

[metazool]

As it stands, streamlit apps won't deploy in our instance of Posit connect - @albags the status update I got was that the Connect installation is being redeployed from scratch as part of the handover of maintenance to the Support team; either the issue is resolved after that, or we have a different front door to take the request and the logs to.

I've made some really small improvements (got rid of the buggy behaviour where cluster contents were rendering twice when you switched between labels; moved a little more of the code into the cyto_ml package) while working through #32.

We run this on an on-prem VM for demo purposes, I think that's fine. I'd love a UX eye on making it more informative / appealing.