Validate initializer types for transpose optimizer (#32647)
This pull request improves the robustness and correctness of the
transpose optimization logic in ONNX Runtime by adding stricter type
checks and handling unexpected or unsupported tensor data types
gracefully. It also expands the test suite to cover these new behaviors,
ensuring that optimizations are only applied when input types are as
expected.
**Key changes include:**
### Type Safety and Robustness
* Updated utility functions like `DataInt64`, `DataInt32`, and
`TensorIntData` to return `std::optional` and check tensor data types
before accessing data, returning `std::nullopt` if the type does not
match expectations. This prevents undefined behavior when tensor types
are incorrect.
[[1]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL34-R49)
[[2]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdR2154-R2181)
* Added explicit data type checks in optimization handlers (e.g., for
Reduce, Slice, Reshape, and Tile ops) to ensure that only supported
types (typically `INT32` or `INT64`) are processed; handlers now bail
out early if an unexpected type is encountered.
[[1]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL1951-R1963)
[[2]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL2199-R2224)
[[3]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL2645-R2684)
### Defensive Programming in Optimization Handlers
* Modified all usage of data extraction helpers to properly handle and
propagate `std::nullopt` cases, ensuring that optimizations are not
applied when the input constant tensors are of the wrong type or
missing.
[[1]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL1977-R1992)
[[2]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL2210-R2239)
[[3]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL2223-R2262)
[[4]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL2306-R2353)
[[5]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdR2572-R2577)
[[6]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL2554-R2592)
### Testing Improvements
* Added utility `ReplaceInitializerWithInt8` to allow tests to inject
initializers with unsupported data types, simulating error cases.
* Added new tests (`TestSliceUnexpectedAxesTypeNoOpt`,
`TestReshapeUnexpectedShapeTypeNoOpt`) that verify the optimizer does
not apply transformations when tensor data types are not as expected,
increasing test coverage for edge cases.
[[1]](diffhunk://#diff-4d8d2dcea7a3e4f5d77f7013252af876e48fc00058ee5b3d4ae6eb0b02b6bbc1R2067-R2102)
[[2]](diffhunk://#diff-4d8d2dcea7a3e4f5d77f7013252af876e48fc00058ee5b3d4ae6eb0b02b6bbc1R4392-R4432)
These changes collectively ensure that the transpose optimizer only
modifies the graph when it is safe and valid to do so, and that it
gracefully skips optimization when encountering unsupported or
unexpected tensor types.
**References:**
[[1]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL34-R49)
[[2]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL1951-R1963)
[[3]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL1977-R1992)
[[4]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdR2154-R2181)
[[5]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL2199-R2224)
[[6]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL2210-R2239)
[[7]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL2223-R2262)
[[8]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL2306-R2353)
[[9]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdR2572-R2577)
[[10]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL2554-R2592)
[[11]](diffhunk://#diff-add01cb10f69ee3e3272b7508ee68ebb98dbbd2060b1385efa7f6f1c1bbe78bdL2645-R2684)
[[12]](diffhunk://#diff-4d8d2dcea7a3e4f5d77f7013252af876e48fc00058ee5b3d4ae6eb0b02b6bbc1R47-R59)
[[13]](diffhunk://#diff-4d8d2dcea7a3e4f5d77f7013252af876e48fc00058ee5b3d4ae6eb0b02b6bbc1R2067-R2102)
[[14]](diffhunk://#diff-4d8d2dcea7a3e4f5d77f7013252af876e48fc00058ee5b3d4ae6eb0b02b6bbc1R4392-R4432)
