Validate initializers before shape propagation (#32653)
This pull request introduces a validation step to ensure that embedded
tensor initializers have data sizes matching their declared shapes
before shape propagation occurs. Additionally, a new unit test is added
to verify that malformed initializers are correctly rejected. These
changes help prevent silent errors during shape inference caused by
invalid initializer data.
Validation improvements for tensor initializers:
* Added a check in `graph.cc` to validate the data size and shape of
embedded (non-external) tensor initializers using
`utils::ValidateEmbeddedTensorProtoDataSizeAndShape` before using them
in shape propagation.
Testing for malformed initializers:
* Introduced a new test in `shape_inference_test.cc`
(`ShapeInferenceDataPropagationTest.RejectsMalformedInitializerBeforeUnpacking`)
to verify that the graph resolution fails with an appropriate error
message when an initializer's data field count does not match the
expected shape.
