Alternative backends

[decahedron1]

This PR adds runtimes other than the ONNX Runtime.

There are plenty of alternative ONNX inference engines for Rust that each provide their own unique qualities:

• candle boasts impressive performance and GPU acceleration.
• tract is a battle-tested pure-Rust inference engine with excellent operator support.
• wonnx focuses on broader GPU support is designed for the web.

With the removal of wasm32-unknown-unknown support from ort, and the Sisyphean task of getting the damn thing to link, it's clear that ONNX Runtime isn't always the best choice. Most often, though, it is the best choice for one platform, but not another. Applications wishing to target CUDA on desktop and WebGPU on web would need to have 2 different code paths using ort and wonnx. Adding support for wonnx & others directly in ort would mean that developers only need to use the ort API to target both backends, and only a single line of code is required to switch between them, e.g. ort::set_api(ort_candle::api());

Status

• ort-candle
  • PoC
  • Tensor values
  • Session creation
  • Inference
  • Alternative 'execution providers' - appears to currently be unsupported by candle-onnx?
  • IoBinding
• ort-tract
  • PoC
  • Tensor values
  • Session creation
  • Inference
• ort-wonnx

_{p.s., sponsorships allow me to spend more time on this PR =)}