v0.2.4

This release improves TimeSeriesImputer performance by combining two implementation wins in the multivariate imputation path.

The first change reduces repeated data movement during imputation. Instead of repeatedly slicing the full lagged matrix for every missingness pattern, the imputer now builds sampled local train and prediction matrices once per target column and reuses them throughout pattern processing.

The second change speeds up the regression step used during imputation. FastRidge.fit now computes the centered normal equations directly, avoiding repeated allocation of centered X and y arrays for each fitted model.