Replace ONNX Runtime inference with PyTorch

[Steffenhir]

Summary

• add a torch-based inference helper that converts ONNX models via onnx2torch and caches them per device
• update deconvolution, denoising, and background extraction to run through the new torch pipeline and select the best available device
• drop the onnxruntime provider helper and add the required torch/onnx dependencies
• update the build workflow to install torch-directml on Windows and drop the onnxruntime-specific packages

Testing

• pytest tests/test_AstroImage.py

---

https://chatgpt.com/codex/tasks/task_e_68e0e878a98c8326a90d26da16e7d0cd

[Domdron]

Would this enable hardware acceleration for denoising on Apple Silicon?