docs: a how-to-guide for using unconditional flows to detect model misspecification

[janfb]

To detect a potential misspecification / out of distribution problem of a real-world x_o, one could train an unconditional flow on the training data x to learn q(x). Given a new x_o one could then evaluate it under q(x) and run common out-of-distribution checks (e.g., permutation tests, or just a cutoff on the log_prob) to detect a model misspecification.

To fix this issue, write a how-to-guide using a simple example.

Depends on #1432

[plcrodrigues]

PR #1470 implements the unconditional flow training. I would like now to write a simple example to illustrate model misspec. Except if @coschroeder is already doing this?

[coschroeder]

here is a simple 2d Gauss example https://github.com/coschroeder/sbi/blob/model_misspecification/sbi/diagnostics/misspecification_toy_example.ipynb

[janfb]

and there is a how-to-guide here: https://sbi.readthedocs.io/en/latest/how_to_guide/18_model_misspecification.html