ENH: Refactor sampling in Variational Inference methods

[zaxtax]

Context for the issue:

Right now in pymc the sample method on Approximation store the model that was used for training. This makes it harder to use variational inference with the posterior predictive sampling and other parts of the Bayesian workflow.

It would be nice if modelcontext was used here instead.

Also, it's probably time to just generate the InferenceData object directly

[RishavIITK22]

@zaxtax Hi!, I am quite new to the world of open source contribution and would like to take up the issue, having prior experience in Probabilistic Machine Learning.