[Feature Request] Outsource Local Inference into its own dedicated container to reduce the main image size

Description The current docker image is ~2.46 GB extracted on disk. Compared to the v2 images (~1.42 GB), this is an increase of roughly 1 GB. Most of that seems to be the new AI feature which can run locally inside the container or remotely via Ollama or an OpenAI compatible API. Especially the local inference part seems to use ~810 MB (checked via image layers). Every installation pays for it, regardless if It's used or not, even though PAPERLESS_AI_ENABLED defaults to false. The problem I face The increase in storage needed generally hurts small scale deployments on e.g. cheap VPS offerings or older hardware like laptops (mine has a 128 GB SSD and a new one with the current marked is not an option). Suggestion: Drop the local inference backend from the main image and run inference as a separate container that Paperless talks to via API. This would reflect the way Immich handles AI workloads and has the additional benefit of giving users a first party possibility to …