1. Port and integrate symbolic shape inference from ONNX Runtime as an analysis pass / feature
2. Create a mechanism in the IR to store sympy expressions. (e.g. an expr field in SymbolicDim?)
3. Allow turning the model info static shapes. (re-evaluation? Direct evaluation?)
4. Allow extension points.

From @jambayk: one current use case is splitting an onnx model into multiple models where we need the shapes for the io connecting the components. another is fixing shape from dynamic to static