Overwrite weights (initializers) with fixed data or random data

[zhenhuaw-me]

Bert series ONNX models are very large (x GB) thus not easy to share the real file. We can improve this process by overwriting the weights (initializers)

• It can be fixed data (e.g. all 0.1 or other value specified), thus the model can be compressed.
• After sharing, we can recover with numpy style random numbers.

This can only be used as a sharing method, the generated model are not useful when evaluate accuracy.

For better usage:

• Annotation will be added when writing fixed data, thus when re-random we can detect automatically.
• The tensors can be specified with names or size.
• Only works for FP32/FP16.
• 0 removed.