Fix two breaking CI failures: reshape-reshape rule test and logsumexp float16 tolerance

[Copilot]

Two unrelated CI failures: ORT-nightly incorrectly folds consecutive Reshape ops with allowzero=1, causing a shape mismatch between the original and rewritten models; torch-nightly exposes float16 precision gaps in logsumexp that exceed default tolerances.

Changes

• _basic_rules_test.py — test_reshape_reshape_dynamic_rule: switch numerical comparison to use_reference=True (ONNX reference evaluator) instead of ORT. ORT's graph optimizer incorrectly folds Reshape(Reshape(x, s1), s2, allowzero=1) when the intermediate has zero-sized dimensions, producing wrong output shapes ((1,0,6,3) vs the correct (1,0,0,3)). The rewrite rule itself is semantically correct; ORT is the broken oracle.

• ops_test_data.py — add tolerance={torch.float16: (2e-2, 1e-4)} to the logsumexp entry. Observed float16 drift in torch-nightly: absolute Δ ≈ 2.3e-5 (limit 1e-5), relative Δ ≈ 2.3e-3 (limit 1e-3).

[justinchuby]

@copilot create a PR on onnxruntime to fix the reshape bug

[Copilot]

@copilot create a PR on onnxruntime to fix the reshape bug

I'm unable to create PRs on external repositories — I can only push to the current microsoft/onnxscript repo. To file the fix in onnxruntime, someone would need to:

1. Locate the graph optimizer pass that folds consecutive Reshape ops (likely in onnxruntime/core/optimizer/reshape_fusion.cc or a similar file).
2. Add a guard: if either Reshape node has allowzero=1 and any of the shape values are 0, the fold should be skipped, because the semantic meaning of 0 changes depending on the context (copy-from-input vs literal zero).
3. Add a test case covering Reshape(Reshape(x, shape_with_zeros, allowzero=0), shape_with_zeros, allowzero=1) with a zero-sized dimension to verify the correct output shape is produced.

The repro is: input shape (0, 6, 2) → first Reshape([3, 2, -1], allowzero=0) → second Reshape([0, 0, 3], allowzero=1). ORT currently gives (0, 6, 3) but the correct output is (0, 0, 3).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 72.52%. Comparing base (79c0a99) to head (91bed69).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2907   +/-   ##
=======================================
  Coverage   72.52%   72.52%           
=======================================
  Files         258      258           
  Lines       31480    31480           
  Branches     2972     2972           
=======================================
  Hits        22831    22831           
  Misses       7640     7640           
  Partials     1009     1009           

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