yolox.py was using a rather suspicious way of getting the available ONNX runtime providers, when there is a perfectly good function called onnxruntime.get_available_providers() that it could be using instead.

Also make it warn the user if CUDA is not present since they need to explicitly install a particular version of onnxruntime-gpu for this to work (see Unstructured-IO/unstructured#2506 for details)