[0z5a]

Summary

When a lazily imported constant is renamed, copy its ONNX TensorProto before applying the new name. This prevents export from mutating the initializer in the original input model while keeping values lazy. The existing fast path still reuses the protobuf when its name is unchanged.

Fixes #4913.

Validation

• 288 non-example ONNX GraphSurgeon Python tests passed and one was skipped on CPU with ONNX 1.20.1. A regression test checks that both the original and exported models pass ONNX validation, the original bytes remain unchanged, and the value stays lazy; it fails on the original implementation.
• The separate example tests were excluded because their subprocesses replace the environment and use a system Python without NumPy in this setup. The docs build and TensorRT sample tests were not run.

Draft while the issue awaits maintainer review under the contribution workflow.