Integration of ONNX runtime

[fortmeier]

To be able to do postprocessing as
** DLSS (deep learning super sampling)
** upsampling

or other techniques as
** latent space decoding

or even just using klartraum as an inference engine, a suitable format representing neural network architecture and tensor contents should be loadable and processable.
The most promising format is ONNX, as it is specifically designed for this purpose and has a nice binary format defined with protobuf.

[fortmeier]

merged first version capable of computing a basic onnx model