Validate dynamic pads input access (#32654)
This pull request strengthens input validation and error handling for
the `ConvTransposeWithDynamicPads` operator. The main focus is on
ensuring that the required dynamic pads tensor is always provided, both
at schema and runtime, and that missing or malformed inputs are handled
gracefully. Additionally, a new test is added to verify the operator's
behavior when the pads input is missing.
**Input validation and schema enforcement:**
* The operator schema for `ConvTransposeWithDynamicPads` is updated to
make the `Pads` input required instead of optional, ensuring that all
usage must provide this tensor.
* The shape inference function now immediately returns if fewer than
three inputs are provided, preventing access to missing inputs.
* In the operator implementation, a check is added to explicitly return
an error if the `Pads` tensor is missing when dynamic padding is
expected.
* The `getInputData` method in `InferenceContextImpl` now safely returns
`nullptr` if the requested input index is out of bounds, preventing
potential crashes.
**Testing improvements:**
* A new unit test, `ConvTransposeWithDynamicPads_MissingPadsRejected`,
is added to verify that the operator correctly rejects models where the
required `Pads` input is missing.
**Test infrastructure:**
* Includes necessary headers in the test file to support the new test
case.
