The backend now knows how big images are, so the frontend can size images correctly before they're loaded.
This should significantly improve the UI on slow connections.
Also fix bug where videos weren't erased from the index properly.
I decided I wanted to integrate the experimental OCR thing better, so I rewrote in Go and also integrated the thumbnailer.
However, Go is a bad langauge and I only used it out of spite.
It turned out to have a very hard-to-fix memory leak due to some unclear interaction between libvips and both sets of bindings I tried, so I had Claude-3 transpile it to Rust then spent a while fixing the several mistakes it made and making tweaks.
The new Rust version works, although I need to actually do something with the OCR data and make the index queryable concurrently.
