[timeseries] Do not impute covariates

Description

Currently the unified preprocessing pipeline for time series imputes the covariates for all items. This is problematic in cases where dynamic covariates contain a lot of missing values. A better option might be to delegate imputation to each underlying model since some models (like Chronos-2) can natively handle NaNs in covariates.