AOTI as a delegate for CUDA, ROCm and MPS support

[larryliu0820]

🚀 The feature, motivation and pitch

Prototype the idea of using AOTI as a delegate in ExecuTorch, figure out the e2e flow to export and run.

Target is to use it on a real world model (such as MV3 or even LLM).

Alternatives

No response

Additional context

No response

RFC (Optional)

No response