Add torch resolution to old plugin

[joerunde]

Feature description

For the 2.0 release, we should consider enabling real dependency resolution for torch

Motivation and context

Given the work to validate the torch_sendnn runtime stack on newer versions of torch, and the confusion caused to users by not including torch in dependency resolution and instead using the system torch install from a base image, we should probably declare torch as a real dependency and resolve it along with our other dependencies.

The locked version can stay at whatever specific torch version we've validated the runtime stack on, and we'll have helpful build failures to remind us when we need to revalidate with a newer torch version. This would be preferable than having our users complain when their builds start failing.

Proposed solution

We can pull over the pyproject.toml snippets from vllm-spyre-next that:

• Force torch to install the cpu flavor
• Add VLLM_TARGET_DEVICE to the builds to ensure we don't compile other device kernels

and pin the torch version to 2.10 for the 2.0 release

Checklist

• I have searched for similar feature requests