p-outlier parameter should consistently be definable as regression

[AlexanderFengler]

We are currently treating p-outlier in a special way, and it implies that this parameter can be estimated, but only globally not via a regression. I don't think there is anything that absolutely forces us to do that, it is rather a legacy decision that wasn't revisited since a long time.

Instead, we should allow p-outlier to be included in the config dict, and if included, it should be treated like any other parameter.

In addition to a bit of extra logic on identifying the presence of p-outlier in the include argument, this will likely need a bit of surgery on our likelihood constructors to respect entry of p-outlier as a vector parameter.