Fix/min_train_samples and min_train_series_length are properly computed

[madtoinou]

Checklist before merging this PR:

• Mentioned all issues that this PR fixes or addresses.
• Summarized the updates of this PR under Summary.
• Added an entry under Unreleased in the Changelog.

Fixes fixes #2264, fixes #1823.

Summary

• min_train_series_length and min_train_samples are now properly depending on each other
• min_train_samples attribute was relaxed for models such as ARIMA or LinearRegression, it is now the value under which, the model raises an error
• historical_forecasts() now rely on min_train_series_length to assess how many steps must be set aside when retrain=True and train_length=None.
• added min_train_past_covariates_series_length and min_train_future_covariates_series_length attributes, they assume that the series are optimally aligned with the target

Other Information

• added tests for the historical forecasts utils functions
• improved parametrization of historical forecasts tests