Stream board generation via SSE.

Accepts application/json (legacy, text-only) or multipart/form-data (new, supports any combination of audio, image, document, and feedback fields). When audio is provided, the stream emits transcribing → transcript before board partials. When an image or PDF is provided, the stream emits analyzing immediately so the UI can show progress before the agent starts.

Delivers partial board fields progressively as the agent generates them. If the client disconnects mid-stream on a text-only turn, generation completes in the background and a push notification is sent. Multimodal turns skip recovery (the upload would have to be redone) and surface the disconnect to the client instead.