CUDA delegate proof of concept

[larryliu0820]

🚀 The feature, motivation and pitch

Make ExecuTorch work on CUDA by integrating AOTI.

Use Voxtral (https://huggingface.co/mistralai/Voxtral-Mini-3B-2507) to verify that the POC works.

Benchmark performance and compare with torchnative (https://github.com/meta-pytorch/torchnative/pull/112/files)

Alternatives

No response

Additional context

No response

RFC (Optional)

No response