[ENH] outlier detection based on probabilitsic regressors

[fkiraly]

There are a few reduction strategies from outlier or anomaly detection to probabilistic regression.

We could support these in skpro, with a pyod compatible interface. This would in turn allow to use the resulting pyod compatible estimators in sktime anomaly and changepoint detectors.

Reducers we could implement:

• quantile based: outlier is extremity of predictive quantile
• density based: outlier is log-pdf of observed value
• loss based: outlier is predictive loss
  • the above two can be seen as special cases, for log-loss or constraint violation

The above being applied to conditional or unconditional distribution estimates.