[BUG] RegressionModel historical forecasts with specific encoder lags

[dennisbader]

Describe the bug
There is a bug when calling historical_forecasts on a RegressionModel that uses encoders with a non-divisible frequency RegressionModels and non-overlapping maximum lag with the target lags.

To Reproduce

from darts.models import LinearRegressionModel
from darts.datasets import AirPassengersDataset

model = LinearRegressionModel(
    lags=3,
    lags_past_covariates=[-3, -2],
    add_encoders={
        "cyclic": {"past": ["month"]},
    }
)

series = AirPassengersDataset().load()
hc = model.historical_forecasts(series=series)

Expected behavior
Should not raise an error. The error

System (please complete the following information):

• Python version: [e.g. 3.10]
• darts version [master]

Additional context
The error can be avoided by passing series also as past_covariates to historical forecasts. Reason is that encoders are generated on the fly with minimum time span requirements -> we have to calculate the expected start index of covariates because it is not available in the time index generated by the encoders.