Simplify the logic of _add_static_covariate in regression models

[eliane-maalouf]

After PR#1399 it would be possible to add static covariates inside of create_lagged_data after each 'block' has been formed.
The current limitation for this is that the implementation tests for the number of features that the regression model was trained with in order to compare it to the current available features at inference time (i.e., to account for the case where the series used at inference time does not have static covariates while the training ones did).
To avoid the latest, one could for example store the static covariates map (that was built at training time) as an attribute of the model and have access to it at inference time.