[Fix][Relax][ONNX] Support runtime shape tensors in Reshape

[javierdejesusda]

The Relax ONNX importer handled Reshape shape inputs that are constants or Shape outputs. When the shape is a runtime int64 tensor, such as a graph input, the converter passed the tensor to relax.op.reshape, which fails with TypeError: Reshape requires the input new shape to be Shape. However, the given one is relax.TensorType.

For a runtime shape, the converter now builds the target shape with Relax tensor ops and converts it to a ShapeExpr with _tensor_to_shape_expr, the way the dynamic-axes path of Unsqueeze does. With allowzero=0 a 0 entry copies the input dimension at the same index, and a -1 entry is the number of input elements divided by the product of the other entries, so neither reaches relax.op.reshape as a literal. The copy index is clamped to the last input dimension so take stays in bounds, and a scalar input skips the copy. With allowzero=1 a literal 0 stays 0. The constant and ShapeExpr paths are unchanged.

The new tests make the shape a graph input and compare the imported model with onnxruntime for the shape from the issue, -1, 0, a rank increase and decrease, a scalar input, a symbolic batch dimension, and allowzero=1 including a zero-size input. One more test checks the error for an unknown shape length, and one checks that a Shape output of an unknown-rank input still imports. The 16 runtime-shape tests fail on main with the TypeError above.

Fixes #20174

Related to #17892, which reports the same error on v0.20.0.

Tests:

• pytest -q -n 6 tests/python/relax/test_frontend_onnx.py: 633 passed, 7 skipped, 4 xfailed
• pre-commit run --files on both changed files: passed