Support Different backend for torchmetric.

[eugene123tw]

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

The torchmetrics.MeanAveragePrecision module now supports the faster-coco-eval backend, which is significantly faster than the default pycocotools backend.

Performance Summary for 5000 Images:

Metric	faster-coco-eval (s)	pycocotools (s)	Speedup (×)
bbox	5.812	22.72	3.91×
segm	7.413	24.434	3.30×

---

Describe the solution you'd like to propose

Support for configurable backends should be added to the relevant recipes, allowing users to choose between faster-coco-eval and pycocotools.

---

Describe alternatives you've considered

• Continuing with the default pycocotools backend despite its slower performance.
• Using a custom integration for faster-coco-eval outside of recipes.

---

Additional Context

• TorchMetrics Documentation: Mean Average Precision
• faster-coco-eval GitHub Repository