ciflow/xpu/186087: [XPU] Keep max_pool2d_with_indices_backward fallback for performance

Per review from jianyizh and guangyey: removing the is_xpu fallback and

re-enabling the scatter_add decomposition causes atomic contention

performance regression with overlapping pooling windows. Revert the

decomposition removal and update the comment to document the performance

rationale (not precision — the kernel fix from torch-xpu-ops #3765 is

landed and pinned).

The fallback stays for performance. Tests reverted accordingly:

• test_torchinductor_codegen_dynamic_shapes: keep TestFailure entries
  
  (no Triton kernel generated while fallback is active)


• test_torchinductor_opinfo: keep inductor_one_sample skip

Co-authored-by: Copilot [email protected]