trunk/36f2cea3c3440f40ac205d70ed4e0a0ac5d5ce90: [XPU] Unskip test_float16_reduction_with_int_output (#185199)

Summary

Removes the @skipIfXpu decorator from CudaReproTests::test_float16_reduction_with_int_output in test/inductor/test_cuda_repro.py. The skip referenced intel/torch-xpu-ops#3006, where Triton codegen for argmax over float16 inputs was emitting a spurious .to(tl.float16) downcast on the int64 index result. That bug was fixed upstream by #174321 ([inductor] Fix incorrect Triton reduction output dtype), which is already in main, so the XPU skip is no longer needed.

Verification

Verified on Intel Data Center GPU Max 1550 (PVC) with torch==2.13.0.dev20260525+xpu:

source /home/gta/intel/oneapi/setvars.sh

export LD_PRELOAD=/home/gta/intel/oneapi/ccl/2022.0/lib/libccl.so.1

export PYTORCH_TEST_WITH_SLOW=1

pytest -v test/inductor/test_cuda_repro.py \

  -k test_float16_reduction_with_int_output -rs

Result:

test/inductor/test_cuda_repro.py::CudaReproTests::test_float16_reduction_with_int_output PASSED [16.3236s] [100%]

====================== 1 passed, 110 deselected in 26.62s ======================

Fixes intel/torch-xpu-ops#3006.

Authored by Claude (opencode).

Pull Request resolved: #185199

Approved by: https://github.com/zou3519