fix(pattern): skip ONNX validation when external data is unloaded

[xieofxie]

Summary

• Fixes a regression from Skip loading external weight data during static analysis. #379 where models loaded with load_external_data=False caused PatternMatcher to log WARNING Model validation failed for pattern matching and skip all pattern detection entirely.
• Root cause: check_onnx_model called onnx.checker.check_model on a proto whose tensors had data_location=EXTERNAL but no raw_data; the checker couldn't find the sidecar .data file and raised ValidationError.
• Adds skip_if_unloaded_external_data: bool = False to check_onnx_model. Default is False so existing callers are unaffected; PatternMatcher opts in explicitly.
• _build_lookups skips numpy_helper.to_array for unloaded external initializers and records their names in _external_unloaded_names.
• _check_constant_constraints skips the value comparison for those names so topology-based pattern matching (Gelu, LayerNorm, etc.) still works without tensor data.
• Adds has_unloaded_external_data utility exported from winml.modelkit.onnx.
• New tests/unit/onnx/test_onnx_utils.py and TestPatternMatchingWithUnloadedExternalData.

[xieofxie]

See warnings like WARNING PatternMatchingValidator: Pattern matching failed - Model validation failed during pattern matching initialization

[xieofxie]

https://github.com/microsoft/WinML-ModelKit/pull/410/changes#diff-da3e2286a2625431d6c74f354f5db3470402b22816d4fc9197268d1c2ff44d23