Implement ONNX GridSample operator with spatial interpolation modes

[HudsonGraeme]

Introduce GridSample supporting bilinear, nearest, and bicubic interpolation with zeros, border, and reflection padding modes.

[kali]

Thanks for the contribution. I had a look, it does look good, I'll have a few remarks/questions later. Let's see what the CI thinks first. We'll also run a "Full" CI on your branch to get coverage across ONNX versions.

[kali]

Ok so that's one of the nice-to-have things that I wanted to talk about actually. We try to make it so ONNX models are serializable and deserializable to tract variant of NNEF, and this is what this CI tests. It is useful for fast startup time with small or mid-size models on small devices, but it may be less relevant for this operators, but...

So this would require:

• move nearly everything to a new module in onnx-opl. InferenceOp impl and onnx attribute extractors stays where they are.
• add a nnef serialization and deserialization to the new module (i think it's pretty straightforward, we have 3 attributes / fields with simple values)

This is a nice-to-have. If you don't want to go there, I'm prepared to accept a PR without it : you can filter out the nnef-cycle test over this new operator.

[kali]

If I did not fudge the GHA setup the full CI is running there: https://github.com/sonos/tract/actions/runs/22661502716