Abstract image descriptions / recognition to use any model provider

[lalalune]

Is your feature request related to a problem? Please describe.

Right now the options for model provider for image recognition are openai and local with florence 2. The OpenAI is hardcoded to check for the existence of the API key, and then load that, otherwise use the local.

Describe the solution you'd like

Similar to the new model provider abstraction, we want to offer image recognition for whatever the current model is. For example, Claude, Vertex Claude, Cloud Llama, etc.