Support models of 2GB or more in make_dynamic_shape_fixed

[adityasingh2400]

Fixes #16773

onnxruntime.tools.make_dynamic_shape_fixed fails on any model at or above the protobuf message limit. fix_output_shapes calls onnx.shape_inference.infer_shapes, which serializes the model into a single protobuf, so the run dies with "Message onnx.ModelProto exceeds maximum protobuf size of 2GB" before anything is written. Saving the result has the same problem.

Shape inference now goes through onnx.shape_inference.infer_shapes_path for models of that size. It keeps the initializer data in a separate file, so the 2GB limit applies to the graph alone. Only the inferred output shapes are read back, so the initializer data stays on disk rather than being loaded a second time. Models below the limit keep taking the existing in-memory path, so nothing changes for them.

This also adds --use_external_data_format so the fixed model can be written with its initializers in a separate file. It turns on automatically for models of 2GB or more, since those cannot be saved any other way.

Verification: added two tests to tools/python/util/test/test_onnx_model_utils.py. One drops PROTOBUF_SIZE_LIMIT to exercise the path based shape inference branch without building a 2GB model, the other drives the command line tool end to end and checks the .data file is written and the shapes are fixed. All 8 tests in that file pass, and both new tests fail against the unmodified code. ruff check and ruff format are clean on the changed files, the two remaining ruff findings in onnx_model_utils.py predate this change.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.