fix(api): keep per-object smoke_type in sequence-to-detection propagation

[MateoLostanlen]

• auto_create_detection_annotations applied smoke_types[0] to every pre-populated detection box, so sequences containing objects of different smoke types (e.g. one wildfire + one industrial) were mislabeled at the detection level.
• Each prediction now inherits the type of the sequences_bbox object whose box matches it: exact coordinate match first (object boxes come from algo_predictions untouched), best IoU as a robustness fallback, and the sequence-wide first type only when a prediction overlaps no object box.
• Adds a unit test on convert_algo_predictions_to_annotation (exact / IoU / no-overlap paths) and an end-to-end regression test (mixed wildfire+industrial sequence); both fail before the fix.

Fixes #142

[Chouffe]

Overall LGTM! Have you verified this empirically with some sequences? Is there a way for me to reproduce it?

[MateoLostanlen]

Yes — verified empirically against the local dev stack (docker compose -f annotation_api/docker-compose-dev.yml up -d --wait --build, API on localhost:5050), with one sequence + one detection carrying two model predictions (left box / right box), then a sequence annotation with a wildfire object on the left box and an industrial object on the right box, processing_stage=annotated to trigger the propagation.

Reading back GET /api/v1/annotations/detections/?sequence_id=...:

On main (bug):

sequence smoke_types: ['wildfire', 'industrial']
  box [0.1, 0.1, 0.2, 0.2] -> smoke_type=wildfire
  box [0.6, 0.6, 0.7, 0.7] -> smoke_type=wildfire   <- industrial object mislabeled

On this branch:

sequence smoke_types: ['wildfire', 'industrial']
  box [0.1, 0.1, 0.2, 0.2] -> smoke_type=wildfire
  box [0.6, 0.6, 0.7, 0.7] -> smoke_type=industrial

Two ways to reproduce:

1. Automated — the end-to-end regression test seeds exactly this scenario through the API:

   cd annotation_api
   make test-specific TEST=tests/endpoints/test_sequence_annotations.py::test_true_positive_sequence_keeps_per_object_smoke_types
   # passes on this branch; to see it fail with the old behavior:
   git checkout main -- src/app/api/api_v1/endpoints/sequence_annotations.py
   make test-specific TEST=tests/endpoints/test_sequence_annotations.py::test_true_positive_sequence_keeps_per_object_smoke_types
   git checkout - -- src/app/api/api_v1/endpoints/sequence_annotations.py

2. Manual — the scenario above via /docs or curl: create a sequence, one detection with two algo_predictions boxes, POST a sequence annotation (stage annotated) with two sequences_bbox objects of different smoke_type each referencing one of the prediction boxes, then read the auto-created detection annotation.

There is also a unit test covering the matching paths (exact coordinates / best-IoU fallback / no-overlap fallback to smoke_types[0]): test_convert_algo_predictions_uses_per_object_smoke_types.