Add Mistral Document AI for document parsing

[r-dh]

This PR is work in progress, and not to be merged as is.

Implementing Mistral requires a lot of architectural decisions, which I want to make explicit and I want to invite discourse.

1. document_processor: MistralOCRConfig | None is concrete if a second processor (Azure Document Intelligence) is added later, this becomes a growing union type and the dispatcher becomes an if/elif chain. A Protocol with a process(doc_path) -> str method would be more extensible. I decided it is acceptable for now with one implementation, but this could rack up tech debt if not properly addressed later.

2. I'm currently discarding the image data entirely. Users get text descriptions but lose the ability to embed or reference original images. I think it makes sense as a default for RAG since we only process text, but it's a silent, non-configurable choice now.

3. Right now, Batch Inference is not implemented. This could greatly reduce costs for large projects. Also, no retry logic on the API call has been added.

Additionally, the tests for Mistral are currently weak and need to be improved before this can be merged.

[emilradix]

@r-dh I marked this as draft, as I understood from your description it is not ready yet to be merged.