Fallback fused RMS normalization fwd/bwd to CPU for functionality

[CuiYifeng]

Fallback _fused_rms_norm and _fused_rms_norm_backward to CPU for XPU backend.

[Copilot]

Pull request overview

This PR adds CPU fallback support for fused RMS normalization operations on the XPU backend. When these operations are called on XPU devices, they will automatically fall back to CPU execution to ensure functionality until native XPU implementations are available.

• Adds _fused_rms_norm and _fused_rms_norm_backward to the XPU fallback list

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[CuiYifeng]

Encountered multiple definition of aoti_torch_xpu__fused_rms_norm when registering _fused_rms_norm in yaml/native/native_functions.yaml. Quickly ensure functionality before solving the multiple definition issue.

[guangyey]

Thanks. I’ve delayed these implementations listed in #1905 and plan to finish them in PT 2.11.

[github-actions]

Performance outliers, please check!

• 🟡 [80%, 90%), may be fluctuations

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
huggingface_float16_training	MT5ForConditionalGeneration	0.914857	0.890362

[CuiYifeng]

Thanks. I’ve delayed these implementations listed in #1905 and plan to finish them in PT 2.11.

Thanks for the reminder. Please note that #1905 also requires modification in stock PyTorch.

[CuiYifeng]

This PR is converted to draft due to some new foundings.