Variationally Informed Parameterization: Recreate VIP class

[ferrine]

https://github.com/pymc-devs/pymc-experimental/blob/d50742d4394fc764cfc0824821b63c235575917c/pymc_experimental/model/transforms/autoreparam.py#L422

So far the VIP transform creates an object with lambdas for manipulation at application time. If any model transform is done later (do, cond), the object is no longer relevant. The idea is to write a helper function as:

class VIP:
    @classmethod
    def from_model(cls, model, var_names: list[str]) -> VIP:
        # Collect lambdas based on name pattern "name" -> f"{name}::lam_logit__"
        ...

[ricardoV94]

Perhaps lambas should be a subclass of shared variable that can be recognized by the type? More robust than by name.

The class could also be replaced by a purely functional approach?

Not sure if we're respecting the class of the shared variables when we convert to fgraph but should be easy to tweak. If not easy, a dummy op could be used to mark them but I don't think we need that