convert DeepFaceLab XSegNet's *.npy weights to onnx file. Inference time is optimized from 500ms to 90ms.

# test on win10 py3.10.14
pip install -r requirements.txt
# remove ./saved_model/.gitkeep for stupid tf

1. put DeepFaceLab XSegNet weights to weights folder, such as weights/XSeg_256.npy.
2. Then run python test_seg.py to generate tensorflow SavedModel format checkpoint file to saved_model directory.
3. convert model to onnx file, python -m tf2onnx.convert --saved-model ./saved_model/ --output xseg.onnx --tag serve.
4. (optinal) install onnxsim pip install onnxsim and run onnxsim ./xseg.onnx ./xseg.sim.onnx.

see test_seg_onnx.py.

Because of Conv2d_transpose requires asymmetric padding which the CUDA EP currently does not support #11312, XSegNet OnnxRuntime Conv2d_transpose layer does not support CudaExcuation.