Fix ONNX node test tensor type conversion (#33150)
### Description
<!-- Describe your changes. -->
Convert serialized ONNX `TensorProto::DataType` values to ABI-stable
`ONNXTensorElementDataType` values before creating tensors in the ONNX
node-test loader. Add the missing `FLOAT8E8M0` conversion.
This fixes the materialized ONNX node tests for `FLOAT8E8M0`, `UINT2`,
and `INT2`, whose serialized values 24-26 differ from the corresponding
C API enum values after #33066.
Validation:
- `lintrunner -a onnxruntime/test/onnx/tensorprotoutils.cc` — passed.
- `cmake --build build/linux_cuda/Release --target onnx_test_runner -j
\"$(nproc)\"` — passed.
- `./onnx_test_runner -e cpu -n cast_FLOAT_to_UINT2
_deps/onnx-src/onnx/backend/test/data/node` — passed (1 test).
- `./onnx_test_runner -e cpu -n cast_FLOAT_to_INT2
_deps/onnx-src/onnx/backend/test/data/node` — passed (1 test).
- `./onnx_test_runner -e cpu -n cast_e8m0_FLOAT_to_FLOAT8E8M0
_deps/onnx-src/onnx/backend/test/data/node` — no failure; the local CPU
build reports the operator as not implemented.
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
The Windows x64 QNN CI on `main` fails both static and shared
configurations because the node-test loader casts ONNX wire type numbers
directly to the C API enum. That cast permutes `FLOAT8E8M0`, `UINT2`,
and `INT2`, producing input/output type mismatches and packed-buffer
size errors in 38 materialized node tests.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
