Fix arange and logit behavior for torch-nightly CI

[Copilot]

Cherry-pick of commit 6c844d3 from copilot/follow-up-bincount-index-put to fix torch-nightly CI failures in aten_arange and aten_logit.

aten_arange / aten_arange_start / aten_arange_start_step

torch-nightly changed the semantics of arange when an integral dtype is requested but one or more inputs are non-integral (e.g. torch.arange(3.1, dtype=torch.int64) now returns 4 elements, not 3). The fix:

• Detects the behavior at import time via _INT64_ARANGE_NON_INTEGRAL_USES_FLOAT_LENGTH
• Routes mixed integral/float inputs through a new helper _arange_integral_dtype_with_non_integral_args that computes the length in float then indexes via Range(0, length, 1) multiplied by step, rather than passing the raw float bounds to op.Range

# torch-nightly: torch.arange(3.1, dtype=torch.int64) → tensor([0, 1, 2, 3])  (len=4)
# previously matched old behavior (len=3); now correctly matches new behavior

aten_logit

The previous two-step clamp (min then max) gave wrong results when eps > 0.5 because 1 - eps < eps. Fixed by applying max (lower bound) first, then min (upper bound), matching torch.clamp order.

# eps=0.6 → one_minus_eps=0.4; old code could clamp self to [0.4,…] then force to 0.6
# new code: clamp to [0.6,…] first, then cap at 0.4 → correct torch behavior

• Fixes Fix CI #2973

[Copilot]

Pull request overview

This PR updates the TorchLib ONNX-function implementations in core.py to match recent torch-nightly behavioral changes that were causing CI failures, specifically around aten::arange* integer-dtype edge cases with non-integral inputs and aten::logit clamping when eps > 0.5.

Changes:

• Add an import-time torch behavior probe and route mixed integral/float arange inputs through a new helper that computes sequence length in float.
• Adjust aten::arange.start_step branching so the “special integral dtype” adjustment applies in the intended cases.
• Fix aten::logit clamping order to match torch.clamp semantics for eps > 0.5.

[codecov]

Codecov Report

❌ Patch coverage is 81.57895% with 7 lines in your changes missing coverage. Please review.
✅ Project coverage is 72.64%. Comparing base (00de62e) to head (d54112a).
⚠️ Report is 13 commits behind head on main.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
onnxscript/function_libs/torch_lib/ops/core.py	81.57%	5 Missing and 2 partials ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2974   +/-   ##
=======================================
  Coverage   72.64%   72.64%           
=======================================
  Files         265      265           
  Lines       32192    32241   +49     
  Branches     3038     3048   +10     
=======================================
+ Hits        23385    23421   +36     
- Misses       7776     7784    +8     
- Partials     1031     1036    +5     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

[Justin Chu (justinchuby)]

Copilot please compute _INT64_ARANGE_NON_INTEGRAL_USES_FLOAT_LENGTH lazily at first use

[Copilot]

Copilot please compute _INT64_ARANGE_NON_INTEGRAL_USES_FLOAT_LENGTH lazily at first use

Done — the probe is now a functools.lru_cached function _int64_arange_non_integral_uses_float_length() evaluated on first use instead of at import time (commit 4d5deb5).