trunk/0467d160b5bc331d23848b3ade51a7eac7570346: Split onehot checks for CPU and accelerators (#179831)

A subtle but significant code change in PyTorch's core library is set to improve performance for developers using Intel's XPU accelerators. The commit, identified as 0467d16 and submitted by a developer, addresses a specific performance regression reported in the torch-xpu-ops GitHub repository (issue #3284). The fix revolves around the 'OneHot' operator, a function commonly used in machine learning for categorical data encoding. Previously, the operator performed boundary validation checks on the CPU for safety, but these checks were intentionally skipped for performance on accelerators like NVIDIA's CUDA and Apple's MPS. Intel's XPU was accidentally omitted from this skip list, forcing costly data transfers back to the CPU and creating a bottleneck.

The developer's solution was elegantly simple: instead of adding 'XPU' to the growing list of exempted accelerators, they flipped the logic. Now, the boundary checks are performed *only* for the CPU, and all accelerators (including XPU, CUDA, MPS, XLA, and PrivateUser1) skip them by default. This aligns with the performance-first philosophy for GPU/accelerator computing, where developers often manage memory and validation explicitly. The pull request was quickly approved by PyTorch maintainers, indicating its correctness and importance for the ecosystem. This optimization is a key example of the continuous, collaborative tuning required to make AI frameworks run efficiently across diverse hardware platforms.