[cuda_xpu_alignment] Fix Triton HOP clone for strided mutated views (#184050)

[Stonepia]

[cuda_xpu_alignment] Fix Triton HOP clone for strided mutated views (pytorch#184050)

Fixes https://github.com/intel-sandbox/agentic_xpu/issues/5

Root Cause: In torch/_higher_order_ops/triton_kernel_wrap.py, the triton_kernel_wrapper_functional_dense and triton_kernel_wrapper_functional_fake_tensor_mode functions use clone_preserve_strides to clone mutated Triton kernel arguments. For strided views (e.g., x[::2]), clone_preserve_strides copies the full storage span, which can read past the bounds of compact Inductor temporaries, producing incorrect values. The upstream fix (commit 9a8a62c) replaces this with a new clone_preserve_strides_for_triton_kernel_wrapper that allocates stride-compatible storage and copies only logical elements via out.copy_(x).

---

Diff stat:

test/inductor/test_triton_kernels.py          | 80 ++++++++++++++++++++++++++-
 torch/_higher_order_ops/triton_kernel_wrap.py | 34 ++++++++++--
 2 files changed, 106 insertions(+), 8 deletions(-)