Weighted anomaly scores

[cgueck895]

Added

• New WeightedRMSEScore anomaly score (weighted_rmse) that applies per-feature weights to
  standardized reconstruction errors before computing the RMSE, enabling feature-specific
  importance in anomaly detection.
• New RMSEScore.standardize() and MahalanobisScore.standardize() helper methods, extracted
  from the respective transform() methods, to share the standardization logic.
• New covariance_method parameter on MahalanobisScore ("auto" | "min_cov_det" |
  "shrinkage"), with support for full-dimensionality (non-PCA) Mahalanobis norm computation on
  mean-centred reconstruction errors.
• New min_cov_det_ratio (default 2.0) and shrinkage_method ("oas" | "ledoit_wolf")
  parameters to control the automatic covariance-estimator selection.
• New documentation page docs/anomaly_scores.rst explaining the available anomaly scores and
  when to use which, recommending RMSE as the default (linked from docs/index.rst).

Changed

• MahalanobisScore now always standardizes reconstruction errors before applying PCA and
  the Minimum Covariance Determinant (MinCovDet) estimation (previously only when scale=True).
• RMSEScore reconstruction errors are always standardized (behavior unchanged from before).
• MahalanobisScore default clean path (pca=False) with covariance_method="auto" now uses
  MinCovDet only when it is stable (n/p >= min_cov_det_ratio and the covariance is full-rank and
  well-conditioned); otherwise it falls back to a shrinkage estimator (Ledoit-Wolf/OAS), keeping
  the score stable for low sample-to-feature ratios and collinear/high-dimensional features.

Deprecated

• RMSEScore(scale=...) and MahalanobisScore(scale=...): the scale parameter no longer has
  any effect and is accepted only for backwards compatibility. Passing it now emits a
  DeprecationWarning. It will be removed in a future release; omit it (standardization is
  always applied).
• MahalanobisScore(pca=True): the PCA-based path is deprecated and will be replaced by the clean
  full-dimensionality Mahalanobis norm implementation. Use pca=False (with covariance_method
  for covariance-estimator control) instead.

Removed

• The anomaly_score.params.scale key from the advanced configuration example
  (docs/examples/advanced_config.yaml) for rmse (and scale from the mahalanobis example
  in the class docstrings).

[roelofsc]

Shouldn't the .standardize() method not be part of the AnomalyScore ABC? Since now all score classes use it.