API process loads onnxruntime/cv2 unnecessarily via worker import chain

Follow-up from #162 (detection-review chunk 1).

Problem

app/main.py imports from app.worker import app as procrastinate_app (needed to open the procrastinate connector in the FastAPI lifespan). app/worker.py in turn does from app.services.smoke_detector import SmokeDetector, and smoke_detector.py imports cv2, onnxruntime, and numpy at module top.

Net effect: the API (uvicorn) process loads onnxruntime + cv2 at startup even though it never runs inference — only the worker process does. This costs ~100 MB RAM and slower startup on the API container for zero benefit.

Suggested fix

Split the lightweight procrastinate App handle out of the heavy task module:

• app/queue.py (or similar): just app = App(connector=PsycopgConnector(...)) — no heavy imports.
• main.py and the enqueue endpoint import the App / task from there.
• Keep the SmokeDetector import (and the @app.task body that uses it) in the worker task module, which only the procrastinate worker process loads.

This keeps onnxruntime/cv2 out of the API process while preserving the "imports at top" convention (no lazy/inline imports).

Notes

• Not a correctness bug — purely a resource/startup-time cost.
• ~15-line refactor touching module structure.

Start with app/main.py, app/worker.py, and app/services/smoke_detector.py to trace the current App and SmokeDetector imports. Then inspect the enqueue endpoint and the task definition before separating the lightweight App handle into app/queue.py or a similar module. Done means the API process no longer loads onnxruntime or cv2, while the worker still imports SmokeDetector and runs its task.