Optimize mlir2onnx.py so that *_final.mlir can be converted to .pb file and visualized in Netron.

The main difference between *.mlir and *_final.mlir is the latter contains information of the subnet and layer group.

Source Code Path:
tpu-mlir/python/tools/mlir2onnx.py

To test:

$ cd tpu-mlir/regression  
$ ./run_model.sh resnet50_tf  
$ mlir2onnx.py -m regression_out/resnet50_tf_bm1684x/resnet50_tf_bm1684x_int8_asym_final.mlir -o resnet50_tf_final.pb

Bonus: 500pt