[BUG] NaiveEnsembleModel does not work with NaiveSeasonal

[turbotimon]

Describe the bug
NaiveEnsembleModel with a NaiveSeasonal with k>3 almost allways gives me an ValueError.
I dont' found a way to prevent this

To Reproduce

from darts import TimeSeries
from darts.models.forecasting.baselines import NaiveMean, NaiveSeasonal, NaiveEnsembleModel
import numpy as np

y = TimeSeries.from_values(np.arange(100))

# NaiveSeasonal alone works fine (no error)
naive_seasonal = NaiveSeasonal(7)
naive_seasonal.fit(y[:50])
naive_seasonal.backtest(y);

# However inside an Ensemble, it gives me a ValueError
naive_ensemble= NaiveEnsembleModel([NaiveMean(), NaiveSeasonal(7)])
naive_ensemble.fit(y[:50])
naive_ensemble.backtest(y);
# ValueError: Train series only contains 3 elements but Naive seasonal model, with K=7 model requires at least 7 entries

# Even if the inputs are dividable by 7 its an ValueError
naive_ensemble= NaiveEnsembleModel([NaiveMean(), NaiveSeasonal(7)])
naive_ensemble.fit(y[:49])
naive_ensemble.backtest(y[:98]);
# ValueError: Train series only contains 3 elements but Naive seasonal model, with K=7 model requires at least 7 entries

Expected behavior
It should work as it works alone

System (please complete the following information):

• Python 3.8.10
• darts 0.23.1

Additional context

• As fare as i figured out, it goes down to the EnsembleModel Base Class
• It is may solvable by taking into account the extreme_lags proberty of each model inside of the ensemble
• However, it is strange to me, that it is also happening when the inputs are dividable by k=7. So i don't even have a workaround at the time! In fact, i didn't found any length of y [0..100] or k>3 so that it works

[JanFidor]

Hi @turbotimon, great find, the error is thrown for ensemble .historical_forecasts() and by extension .backtest() if it has any model with input_chunk_length > 1 or lags looking further than -1. I think I found the cause, when calling historical_forecasts and iterating over historical_series, minimum length is set incorrectly, after digging around with a debugger I found out that the Ensemble's max target lag (in extreme_lags) is set to 1 while it should've taken into account it's models extreme lags. I'll make a quick PR @dennisbader @hrzn @madtoinou