[WIP] experimental: c export of ModelProto in python package

[xadupre]

Description
Many functions serialize a proto into bytes then deserialized in the C++ binding (checked, infershapes, ...) It could be faster by avoiding this step. This PR tries to make ModelProto and other available through pybind11.

Motivation and Context

• Performance

[daquexian]

It is the feature I've been looking forward to for a long time. Serializing and deserializing between c++ and python waste much time in onnx optimizer.

I have another idea to discuss: Why do we stick to protobuf and not switch to another more flexible format like flatbuffers? It is so strange that onnx, an op standard, compromises because of a replaceable storage format. But the fact is we even create a specific mechanism (external data) just to solve the problem of protobuf (cannot handle >2GB data).

cc @onnx/steering-committee @onnx/sig-archinfra

[xadupre]

Thanks for the feedback. The other implementation I tried is #4281. However, depending on how protobuf 4 is integrated into onnx, this option may not be available anymore, see https://developers.google.com/protocol-buffers/docs/news/2022-05-06.