To improve the backward compatibility of torch.onnx.export dynamo=True/False (torchscript-based and torch.export-based), dynamic_axes needs to be converted to dynamic_shapes.

ONNX has a naive approach to convert dynamic_axes to dynamic_shapes.

To avoid the ordering issue (torch.export dynamic_shapes requires None to mark optional inputs), the implementation uses model.signature to tuplify inputs and dynamic_shapes, which results in lower coverage.

After #150583, there should be spaces to improve, where we should be able to unflatten dynamic_axes by following args/kwargs tree structure.