Fix histc(out=) to reject mismatched out dtype instead of silent cast

[N0AHZACH]

Fixes #5237.

_histc_out_xpu computed the histogram in the input dtype via _histc_xpu
and then did resize_output + result.copy_(ret). copy_ casts silently,
so a float64 histogram written into an int64 out= was truncated with no
error. CPU rejects the same call in histogramdd_prepare_out
(aten/src/ATen/native/Histogram.cpp:132), so this was a CPU-vs-XPU
argument-validation divergence.

Changes:

• src/ATen/native/xpu/SummaryOps.cpp: add the CPU-matching TORCH_CHECK
  before resize_output so a rejected call leaves the caller tensor
  untouched. Message text matches CPU verbatim and matches upstream
  Fix histc(out=) dtype validation mismatch between CPU and CUDA pytorch/pytorch#196100 for CUDA.
• test/xpu/xpu_test_utils.py: de-scope tuple device_type=("cuda", "xpu")
  xfails gated on _cuda_xfail_xpu_pass by dropping only xpu from the
  scope. Without this the fixed test reports an unexpected success.
• test/xpu/skip_list_common.py: remove test_out_histc_xpu_float32.
• test/regressions/test_histc.py: regression coverage for mismatched
  out= rejection and matching-dtype success.

This does not depend on pytorch/pytorch#196100 landing and does not
conflict with it.

[N0AHZACH]

@laifenxiawucha could you check this out so that I can get the CI rolling

[N0AHZACH]

@laifenxiawucha can we start the CI on this specific commit - I do see a CI workflow has been run but idk where it ran tho.......been searching for a good 15 mins. Could we start the CI on this specific commit. I'm super sorry if its something stupid from my end

[laifenxiawucha]

Hi @N0AHZACH .No worries at all — this is likely not an issue on your side. Since the PR is from a fork, some CI jobs may need maintainer approval before they can start. That’s probably why you can see a workflow run but not the full CI for this specific commit.

[N0AHZACH]

Thanks @laifenxiawucha. I traced the four failing checks and they are stale, not caused by this change.

The linux-build and windows-build jobs fail before compiling anything (run 35192511198, commit ee3cc7cf):

• Linux: python: can't open file '.../pytorch/.ci/manywheel/build_env_setup.py': [Errno 2] No such file or directory
• Windows: ModuleNotFoundError: No module named 'build_env_setup'

Both come from PyTorch main relocating the wheel build scripts to .ci/wheel/linux and .ci/wheel/windows (pytorch/pytorch#189275). That breakage was fixed on main by #5417 and #5420, and the current head d355f901 already merges them (main at e9e0bfe4), so .github/scripts/build.sh and build_windows.py now use the new paths.

The change itself is already validated on hardware. In run 35050000160 (272fcabb), where the UT actually ran:

• op_regression passed, including both new cases in test/regressions/test_histc.py (test_histc_out_rejects_mismatched_dtype, test_histc_out_matching_dtype_still_works)
• test_out_histc_xpu_float32 ran and passed in op_ut (it is no longer skipped, and no longer reports an unexpected success against the upstream ('cuda', 'xpu') xfail)
• the remaining op_ut entries in that run are unrelated (dynamo / flex-attention / fx / libtorchbind / env), none of them touch histc

Could you approve/start CI on d355f901? The run for that commit (35302042384) is still waiting on maintainer approval. Thanks!