fix(workflows): preserve image metadata on empty Template Matching results - #3038

Summary

When template_matching@v1 finds no matches, its numpy path returns a bare sv.Detections.empty(). The serializer then emits image.width and image.height as null, and the empty result has no parent/root lineage. Use the existing empty_detections_with_image_metadata(image) helper for this branch so zero-match results keep the input image's dimensions and lineage. The matched-result path is unchanged; the tensor-native sibling already carries image metadata on empty results.

Root cause and before/after

• Before: a 160×120 input with zero matches serialized as {"image": {"width": null, "height": null}, "predictions": []}.
• After: the same path serializes as {"image": {"width": 160, "height": 120}, "predictions": []} and retains parent/root metadata on the empty detections object.

What changed

• Return the existing metadata-aware empty detections helper from Template Matching's zero-match branch.
• Add a block-level regression test for zero predictions, serialized dimensions, and parent/root lineage.
• Add the required Workflows Unreleased changelog entry.

Verification

• Before the fix, the new focused test failed on the null dimensions (1 failed, 4 passed).
• After merging current upstream main, uv run --project . --extra test --locked pytest tests/unit_tests/core_steps/classical_cv/test_template_matching.py -q from workflows/: 5 passed, 8 Pydantic deprecation warnings.
• Related serializer, utility, and formatter tests: 368 passed, 84 deprecation warnings.
• Black, isort, and git diff --check: passed.

Scope

This addresses the numpy Template Matching zero-match path. It does not change tensor-native behavior or claim that every downstream Workflows consumer has been exercised. Release timing is up to maintainers.