add YOLOv10-style Equinox detection head for object detection outputs

[bernardev254]

Summary

This PR introduces a new Equinox module YOLOv10Head that implements the prediction head for object detection, inspired by the YOLOv10 architecture.

Changes

• Adds a YOLOv10Head class that maps CNN feature maps to detection outputs.
• Supports outputting bounding box coordinates, objectness scores, and class probabilities.
• Produces output of shape (B, H, W, A * (C + 5)) suitable for post-processing.
• Leaves activation (e.g., sigmoid or softmax) as an optional downstream step.

Motivation

This detection head will integrate with a CNN backbone to form a complete object detection model. It is structured to support modern YOLO-style inference and training, aligning with the overall goal of enabling a performant object detection pipeline within the Equinox + JAX framework.

closes #4