Make ONNXRuntime work with older and newer versions of onnxruntime

[tmadlener]

The current implementation does not compile with onnxruntime@1.17.1. It works in CI, because it picks up v1.10.0 there. This PR should make this work with older and newer versions of the library. The most notable changes are

• FindONNXRuntime now creates an onnxruntime::onnxruntime target (similar to the one that is provided by newer versions of onnxruntime) and this is used throughout the CMake configuration.
• The Cmake configuration first tries to find onnxruntime (i.e. the new version) and only falls back to ONNXRuntime if it can't find that
• The actual c++ implementation should now work with older and newer versions of the API, by switching to non-deprecated functionality and rigging some things on the fly if necessary for older versions.
  • It also moves away from the Experimental namespace and directly goes for the Ort namespace, which should be available for 1.10.0

The last bit required the introduction of an Ort::MemoryInfo that is necessary to create tensors. This now creates the tensors on the CPU, I am not sure if that is going to be an issue or not for running things on GPU. Is running on GPU something that is currently done?

[tmadlener]

@jmcarcell this should make this cherry-pick obsolete: https://github.com/key4hep/key4hep-spack/blob/dc0b2bc4aaad8f3ef5f3c7f62ff5b0277a554eee/.cherry-pick#L37

(Also that threw me for quite some loop ;) )

[kjvbrt]

Thanks @tmadlener :)