TensorRT SegmentationModel fix

[glenn-jocher]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhancements in exporting models with added support for Segmentation and Detection in ONNX and TensorRT formats.

📊 Key Changes

• DetectionModel and SegmentationModel classes added to the list of imports from the models.yolo module.
• Support for exporting SegmentationModel to ONNX with appropriate output names and dynamic axes configuration.
• Dynamic axes handling now considers SegmentationModel and DetectionModel instances differently when exporting to ONNX.
• Adjusted TensorRT code to accommodate multiple output names and dynamic reshaping of inputs and outputs according to the model type.

🎯 Purpose & Impact

• 🎨 Allows for more model types to be exported, specifically those used for segmentation tasks, addressing a wider range of machine learning use cases.
• ⚙️ Improves interoperability by exporting models to formats compatible with different deep learning inference engines such as ONNX and TensorRT.
• 👥 Users benefit from an expanded set of models that can be exported, potentially increasing the adoption of Ultralytics YOLOv5 for varied applications.

[glenn-jocher]

@AyushExel @Laughing-q this PR adds SegmentationModel TensorRT inference support. I updated CoreML and Paddle inference yesterday to support SegmentationModels.

The only SegmentationModel exports not running inference correctly still are the TF formats.

You can see Segmentation benchmark results as part of the daily CI:
https://github.com/ultralytics/yolov5/actions/runs/3077032938/jobs/4971602647