Shape inference support for qonnx symbolic operators

[volcacius]

Setting shape inference is currently broken in upstream PyTorch for custom ONNX operators pytorch/pytorch#81693 . This impact qonnx export, and beyond raising a warning in some use cases lack of shape propagations means that certain other operators (e.g. globalmaxpool) that require correct shape inference are not exported correctly.

[lix19937]

You can ref
https://pytorch.org/docs/1.11/onnx.html#pythonop-symbolic https://github.com/Xilinx/brevitas/blob/master/src/brevitas/export/onnx/standard/function.py