IO overhaul

[decahedron1]

Session inputs/outputs should probably be reworked. The way they work now is a bit messy for my liking, and there is currently no clear path for supporting important features like IOBinding or non-tensor types.

A few things I think should be addressed:

• Allow for preallocating inputs/outputs to improve performance (Allocate outside of run function #37)
• Implement IOBinding (IOBinding #15)
• Support for sequence<T> and map<K, V> types via Vec<T> and HashMap<K, V> (Support for sequence/map types #30)
• Support converting tensors to/from Vec<T> with a given shape (since some applications don't really need the input/output to be a complex tensor, i.e. Silero VAD)

Allocate outside of run function: Discussed in #37

^{Originally posted by stexa April 6, 2023}
Hey!
For the usage in audio processes and in general for the performance it would be really nice if the inputs and outputs could be set and allocated before the actual run() function. Somewhat like suggested here:
nbigaouette/onnxruntime-rs#41

This could be an additional function as well. Might this be something you would be interested in as well?

And thank you, I am very happy that someone is still working on a onnxruntime wrapper in rust, the crate is working out of the box like a charm for me :)

IOBinding: Discussed in #15

^{Originally posted by dzhao January 30, 2023}
Hi, do you plan to support IOBiding for cuda/tensorrt?
https://stackoverflow.com/questions/70740287/onnxruntime-inference-is-way-slower-than-pytorch-on-gpu

This seems a critical feature for gpu serving.