Add a CPU runtime contract test for mixed-width QMoE

[titaiwangms]

Motivation

#744 adds Mobius export support for mixed-width com.microsoft::QMoE, including the target layout:

• FC1/FC3 routed expert weights: INT2
• FC2 routed expert weights: INT4

microsoft/onnxruntime#32697 added the corresponding FC-specific schema attributes and packed-layout validation. CPU, CUDA, and WebGPU intentionally return NOT_IMPLEMENTED for mixed-width execution until projection-aware kernel dispatch is implemented.

The current Mobius CI wheel does not contain the new FC-specific attributes, so #744 cannot yet prove that its exported model reaches the intended runtime boundary rather than failing earlier because of an invalid schema, shape, zero point, or initializer binding.

Scope

Add a CPU negative runtime contract test using an ONNX Runtime build that contains microsoft/onnxruntime#32697.

1. Build or load a tiny Mobius mixed-width QMoE model with FC1/FC3 INT2 and FC2 INT4.
2. Load it with CPUExecutionProvider.
3. Verify the model passes schema and packed-layout validation.
4. Run the model far enough to reach the CPU QMoE execution boundary.
5. Assert that execution fails with the expected mixed-width NOT_IMPLEMENTED status, not an earlier malformed-model error.
6. Record or gate on an ORT version/commit containing the FC-specific QMoE attributes.

Acceptance criteria

• The test exercises the final serialized model and bound initializers produced by Mobius, not only an isolated operator fixture.
• FC1, FC2, and FC3 effective widths and packed tensor shapes are checked through ORT validation.
• A missing attribute, invalid shape, invalid zero-point geometry, or unbound initializer causes the test to fail rather than being accepted as the expected result.
• The only accepted runtime failure is the explicit mixed-width QMoE NOT_IMPLEMENTED boundary from the CPU provider.
• The test is clearly documented as a negative contract test and makes no numerical-parity or execution-support claim.
• Once ORT CPU mixed-width execution is implemented, replace or extend this with positive logits/parity coverage.

Dependencies

• Exporter implementation: Support mixed-width QMoE export #744
• Mobius mixed-QMoE tracker: Support per-layer Olive mixed-precision QMoE export #735
• ORT schema/validation contract: microsoft/onnxruntime#32697
• Positive numerical inference remains blocked on projection-aware CPU QMoE execution.