Normalize list-returning detection models

[CyMule]

Summary

• normalize legacy list results from object-detection models into LayoutElements
• perform normalization at the shared model boundary
• update Detectron coverage and add a page-layout regression test

Why

Some built-in detection models return a list while page layout code expects LayoutElements and reads attributes such as routing. That mismatch can raise AttributeError in a normal inference route.

Compatibility note

The shared object-detection call now consistently returns LayoutElements. Out-of-tree callers that directly consumed a raw list should use .as_list() when they need the prior shape. The draft status leaves room to confirm that boundary contract.

Validation

• uv run --locked --no-sync pytest -q test_unstructured_inference/inference/test_layout.py test_unstructured_inference/models/test_detectron2onnx.py — 59 passed
• Ruff check and format check passed for all changed Python files
• changelog and package versions both set to 1.6.14