Support for YOLO-style / modern object detection models (2D and 3D)

Inizia leggendo monai/apps/detection/retinanet_detector.py e retinanet_network.py, quindi esamina le transform di supporto per il rilevamento, le utility per gli anchor, le metriche e il tutorial RetinaNet esistente. Confronta l’API attuale di DetectorNetwork e le convenzioni per le box con l’ambito 2D e 3D richiesto in stile YOLO. Il completamento dovrebbe includere un ambito del detector concordato, l’integrazione dell’API, la gestione del formato delle box, i test e un tutorial.

Is your feature request related to a problem? Please describe.

MONAI currently ships a single object-detection architecture — RetinaNet —
under monai/apps/detection/ (retinanet_detector.py, retinanet_network.py),
with supporting anchor utilities, COCO-style mAP metrics, and box transforms. This
is a solid anchor-based, two-stage-style foundation and works in both 2D and 3D.

However, the detector zoo stops there. Users who want fast single-stage or
anchor-free detectors most notably the YOLO family — have no native path.
The existing guidance (see #903) is essentially "use MONAI transforms inside your
own YOLO pipeline," which leaves the detector itself, training loop, box-format
handling, and metrics outside MONAI's guarantees. #292 raised YOLO/COCO/Pascal-VOC
box-format support in transforms years ago but there is no dedicated tracking
issue for native modern detectors, and #8519 lists surgical instrument
localization/detection as a target without naming an architecture.

Describe the solution you'd like

Expand monai/apps/detection beyond RetinaNet to include modern detectors, with
YOLO as the flagship because of its strong fit for 2D medical / endoscopy /
surgical-tool and microscopy use cases (real-time inference, anchor-free variants,
mature ecosystem). Concretely:

1. A YOLO-style detector network (e.g. an anchor-free YOLO head) integrated into
   the existing DetectorNetwork / detector API so it reuses MONAI's box
   transforms, anchor/anchor-free utilities, ATSS-style matching, and mAP metrics.
2. Native handling of the YOLO box format (normalized cx, cy, w, h) in the
   detection transforms, alongside the existing corner/CCWH conventions (follow-up
   to #292).
3. A tutorial mirroring the existing RetinaNet LUNA16 / detection tutorial so the
   new detector is a drop-in alternative.
4. (Optional / stretch) Room in the API for other modern detectors such as
   DETR-family or FCOS, so this is an extensible "detector zoo" rather than a
   one-off.

Note: interest in YOLO-inspired 3D detection

MONAI's biggest differentiator over general-purpose CV libraries is first-class
3D support — RetinaNet here already runs on volumetric data. If there is
community/maintainer interest, a YOLO-inspired 3D detector (a single-stage /
anchor-free volumetric detection head operating on 3D feature maps, predicting
6-DoF axis-aligned 3D boxes) would be a genuinely novel and high-value addition:
few libraries offer a fast single-stage 3D detector, and use cases like nodule /
lesion / landmark detection in CT and MR volumes would benefit directly. I'd be
happy to help scope and prototype this if the maintainers see value — please
comment if there's appetite for the 3D direction specifically.

Describe alternatives you've considered

• Continuing to use RetinaNet only (works, but no fast single-stage/anchor-free
  option, and no YOLO ecosystem interop).
• Running an external YOLO (ultralytics etc.) alongside MONAI purely for transforms
  (the current #903 answer) — loses MONAI's metric/box/3D guarantees and
  reproducibility.

Additional context

• Existing detection module: monai/apps/detection/ (RetinaNet).
• Related prior discussions: #292 (box-format transforms), #903 (YOLO usage
  question), #8519 (MICCAI submissions incl. surgical detection/localization).
• Willing to contribute an implementation and tutorial, and specifically to
  prototype the 3D variant if there's interest.