Methods for Detecting Model Misspecification

[gmoss13]

We should implement checks for when the input condition is out of distribution (i.e. the simulator model is misspecified relative to the ground truth data).

Specifically, we can start by implementing MMD checks on the learned embeddings as in https://arxiv.org/abs/2406.03154.

Concretely, after training, when calling posterior.set_x(), we should run the misspecification test, and throw a warning to the user if the check identifies misspecification. We should also add tests where the model is misspecified to see when we can detect it.