Apply upstream Inductor fix for adaptive_avg_pool2d flatten-sum wrong results

[Copilot]

torch.compile(..., backend="inductor") on XPU can return silent wrong results for adaptive_avg_pool2d(x, 7).flatten(1).sum(dim=-1) even though CPU and XPU eager are correct. The failure matches the upstream Inductor contiguous/exact-stride bug fixed in pytorch/pytorch#180898.

• What changes

  • Add upstream PR pytorch/pytorch#180898 to the default patch list applied by /home/runner/work/torch-xpu-ops/torch-xpu-ops/.github/scripts/apply_torch_pr.py.
  • This ensures the stock PyTorch checkout used by torch-xpu-ops CI/build flows picks up the upstream Inductor fix automatically, along with its regression coverage.

• Why this is the right layer

  • The incorrect result is not in the XPU adaptive average pooling kernel in this repository.
  • The bug is in the upstream Inductor lowering/layout logic, so the minimal repo-local fix is to apply the upstream PyTorch patch during the existing PyTorch preparation step.

• Operational impact

  • Existing workflows that clone PyTorch and run apply_torch_pr.py will now include the upstream Inductor fix without any additional wiring.
  • No XPU kernel, YAML, or local operator registration changes are introduced.

• Patched upstream PR

  default=[
      # Additional PRs link if need for CICD tests
      "https://github.com/pytorch/pytorch/pull/180898",
  ]

[github-actions]

Performance outliers, please check!

• 🔴 [-1, 80%), should be regression

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
timm_models_bfloat16_training	vit_base_patch16_siglip_256	0.695962	0.744412
timm_models_bfloat16_training	visformer_small	0.715022	0.754084
timm_models_bfloat16_training	dm_nfnet_f0	0.640517	0.767955
timm_models_bfloat16_training	deit_base_distilled_patch16_224	0.749122	0.775495
timm_models_bfloat16_training	beit_base_patch16_224	0.735098	0.814096
timm_models_bfloat16_training	mobilenetv3_large_100	0.784127	0.823835
timm_models_bfloat16_training	nfnet_l0	0.736359	0.827558
timm_models_bfloat16_training	inception_v3	0.784848	0.836940
timm_models_bfloat16_training	adv_inception_v3	0.787293	0.844838
timm_models_bfloat16_training	mobilevit_s	0.722290	0.960236

• 🟡 [80%, 90%), may be fluctuations

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
timm_models_bfloat16_training	convnextv2_nano.fcmae_ft_in22k_in1k	0.808973	0.815837
timm_models_bfloat16_training	swin_base_patch4_window7_224	0.837380	0.828020
timm_models_bfloat16_training	deit_tiny_patch16_224.fb_in1k	0.865131	0.840690
timm_models_bfloat16_training	repvgg_a2	0.828854	0.846576
timm_models_bfloat16_training	ghostnet_100	0.845650	0.850532
timm_models_bfloat16_training	mobilenetv2_100	0.839038	0.872309
timm_models_bfloat16_training	tf_efficientnet_b0	0.800028	0.907076