Anomaly Detectors Capabilities

[sarusso]

This PR adds the capabilities for the anomaly detectors, which must be implemented in any subclass and that are both enforced and checked at class construction-time.

The capabilities follow the taxonomy reported in the table below:

Capability	Option	Meaning
Mode	Unsupervised	Does not require labeled data or explicit normal/reference data to detect anomalies.
	Semi-supervised	Requires explicit normal/reference data, but no labeled anomalies.
	Weakly-supervised	Requires some labeled anomalies, but not exhaustive labeling.
	Supervised	Requires extensive or exhaustive labeled anomalies.
Streaming	Yes	Can process observations incrementally as they arrive and produce anomaly flags without re-fitting (may include logical/reconstruction delay).
	No	Requires access to the full data required by the model to evaluate anomalies, or requires re-fitting to process new observations.
Context	Point	Requires just a single data point, each timestamp is processed independently, without temporal context.
	Window	Requires a finite local neighborhood of the target series.
	Series	Requires the full history of the target time series, but does not depend on other series.
	Dataset	Requires access to the full dataset, potentially including multiple time series and cross-series statistics.
Granularity	Series	Can only mark entire time series as anomalous or normal.
	Point	Can mark specific timestamps as anomalous.
	Variable	Can label specific variables as anomalous (only for multivariate).
Multivariate	No	Can process only univariate time series.
	Yes	Can process multivariate time series.
	Only	Must process multivariate time series.
Scope	Specific	Model parameters are tied to a single time series or dataset.
	Agnostic	A single fitted model can be applied to unseen time series or datasets.

@clarasaja and @agataben let me know if you see any major issues with the above nomenclature.