Avoid inconsequential warning when using multi-dimensional x

[alvorithm]

Problem

The user is free to supply an embedding net whose output does not reduce x to one dimension (shape (D,)). It is difficult for us to prevent this mismatch, because it requires poking into an already built Sequential architecture, which could turn out to be hacky (but this has to be assessed? see https://github.com/mackelab/sbi/blob/292396b87f9f27806445cb74307f9b58f125e787/sbi/neural_nets/classifier.py#L95 @janfb @michaeldeistler ).

At the moment any sort of multidimensional x will trigger a warning, but the warning will sometimes be idle, i.e. things will just work. We want to prevent idle warnings.

Solution

Do not issue an inconditional warning, but if a dimension mismatch happens at the input of the density estimator proper, catch the exception and include a potential debug hypothesis in the error message ('This could have failed because x has more than one dimension').

Note this has to be handled differently depending on the inference method.

[janfb]

Looks good, thanks for summarising this @Meteore . I will tackle it next week!