Add ExecuTorch support for RF-DETR

## **Description** We would like to add support for **ExecuTorch** to enable efficient, on-device inference for the **RF-DETR** model. Bringing ExecuTorch capabilities to RF-DETR will allow deployment on edge devices (mobile, embedded systems) with optimized performance and reduced memory footprints. ## **Motivation** As on-device AI deployment becomes crucial for real-time applications, exporting RF-DETR models to the ExecuTorch ecosystem will bridge the gap between high-accuracy object detection and edge hardware constraints. This implementation can leverage or align with ongoing ExecuTorch integration efforts within the ecosystem, such as the discussions and implementations highlighted in [react-native-executorch issue #797](https://github.com/software-mansion/react-native-executorch/issues/797). ## **Proposed Solution** * **Export Pipeline:** Implement an export script/pipeline to convert the RF-DETR model into the ExecuTorch ahead-of-time (AOT) compiler format (`.pte`). * **Operator Coverage:** Verify that all specific operators used in RF-DETR (e.g., custom attention mechanisms or specialized bounding box heads) are covered by ExecuTorch's native operator library or define custom operators if needed. * **Validation & Benchmarking:** Provide a basic utility script to validate the numerical accuracy of the exported `.pte` model against the original PyTorch model. ## **Additional Context** * **Reference Issue:** https://github.com/software-mansion/react-native-executorch/issues/797 * If there are specific hardware backends (e.g., CoreML, QNN, XNNPACK) that should be prioritized for testing the RF-DETR deployment, please let us know in the comments.