[Feature Request] Easier installation of onnxruntime-gpu

[gedoensmax]

Describe the feature request

I would like to see an easier installation of the pypi package onnxruntime-gpu. Ideally the package will automatically install:

nvidia-cudnn-cuXX 
nvidia-cublas-cuXX
nvidia-cufft-cuXX
nvidia-cuda-runtime-cuXX
nvidia-cudarand-cuXX
tensor-cuXX

Where XX denotes the depending CUDA version. This should work for Linux and windows equally and heavily ease the entrypoint by not requiring putting all the needed libs on the PATH.

Ideally it would even be possible to install the needed libraries for cuda and/or art only by doing pip install onnxruntime-gpu[cuda]
I am happy to contribute or help testing for this, but I could use someone giving me hints where to look.
@jywu-msft for viz

Describe scenario use case

This will positively impact all python users of ONNX Runtime that use GPU acceleration. Hopefully also reducing library installation issues on GitHub significantly.

[chilo-ms]

@snnn
Could you help review this feature request and provide some guidance? Thanks!
This can really help users run ORT through package without having to manually do other env setting or dependency downloading.

[snnn]

Dup with #19350 .