Set keep_size=True in v4

[OriolAbril]

By default, the output of sample_posterior_predictive and fast_sample_posterior_predictive returns an array with the chain and draw dimensions flattened. This makes conversion to InferenceData more complicated. I propose changing the default to keep_size=True, simplify the code in to_inferencedata and do not support conversion to inferencedata without keep_dims. Users would still be able to get the flattened samples, but to use "on their own". cc @rpgoldman

[brandonwillard]

Sounds fine to me.

[rpgoldman]

I’d suggest you make keep_size default to the same value as the value of return_inference_data to avoid unnecessary breakage to user code. If you do that, then old user code that uses the old return data structures instead of InferenceData will behave as before.