Reactor reinstalling older onnx and onnxruntime-gpu and then failing to work

[levicki]

I am a bit unortodox here in my AUTOMATIC1111 setup in that I am using Torch 2.5.1+cu124, and I have custom compiled xformers using same CUDA version for my RTX 4090.

Unless I install matching stuff manually for this extension and then add --skip-install to command line arguments, it will reinstall some onnx 1.16.1 (even if I installed 1.17.0) and onnxruntime-gpu 1.17 which from this table is clear won't work with newer torch than 2.4.0 because new torch uses CUDNN 9, not 8 which is required by older onnxruntime-gpu:

It would be nice if install script checked the existing torch version so it doesn't try to install onnxruntime-gpu with mismatched CUDNN in the future.