[Fix][Relax] Honor ONNX Reshape zero semantics

[tandede]

ONNX Reshape uses zero entries to copy the corresponding input dimension by default. The Relax importer currently constant-folds those shapes through NumPy without applying that rule, so a valid shape such as [0, 3] fails for a (2, 3) tensor. Conversely, when allowzero=1, passing the zero through relax.op.reshape invokes Relax's own zero-copy convention instead of preserving the literal zero dimension.

This change normalizes copied dimensions before constant folding when allowzero is disabled. When it is enabled, the shape tensor is materialized as a symbolic ShapeExpr, preserving literal zero dimensions without applying Relax's zero-copy shortcut.

The regression tests cover both the default all-constant path and an executable allowzero=1 model with a (0, 2) output.

Fixes #20151

Testing:

pre-commit run --files python/tvm/relax/frontend/onnx/onnx_frontend.py tests/python/relax/test_frontend_onnx.py
pytest -q tests/python/relax/test_frontend_onnx.py -k reshape
pytest -q -n 4 tests/python/relax/test_frontend_onnx.py -k 'not test_clip_v13'

The first two commands pass, and the broad run reports 493 passed, 9 skipped, and 4 xfailed. Running the entire file also reports three test_clip_v13 failures that reproduce unchanged on a clean main worktree in the same environment.

[tlopex]

allowzero only changes zero semantics, but this branch also bypasses -1 inference. A valid non-constant (3, 4) input with shape [-1, 2] and allowzero=1 reaches runtime with a literal -1 and fails with std::bad_alloc. Please keep the normal reshape path when no literal zero is present and add a regression test.

[tandede]

allowzero only changes zero semantics, but this branch also bypasses -1 inference. A valid non-constant (3, 4) input with shape [-1, 2] and allowzero=1 reaches runtime with a literal -1 and fails with std::bad_alloc. Please keep the normal reshape path when no literal zero is present and add a regression test.

Thanks for catching this. You’re right that allowzero should only change the interpretation when the target shape actually contains a literal zero; it should not bypass normal -1 inference.

I narrowed the special path to constant target shapes containing zero. With allowzero=1 and [-1, 2], a non-constant (3, 4) input now follows the normal reshape path and produces (6, 2).

I added the requested regression test and pushed the update in d20b94e. All five Reshape-related tests and the changed-file pre-commit checks pass.