[fx-importer] the importer is slow

[qianfenchi]

Seems like converting a torch model to mlir using fx is really slow, especially for large models.
For example, we convert Qwen3-8B model to mlir with our customized dialect, using torch it takes more than 30mins (fx importer itself takes 20+mins), while using onnx it only takes 2~3mins.

Is this mainly caused by using mlir python bindings?
With onnx, we load the model and create ops in cpp.

Anyone has the same issue? How did you solve that? tks.