[BUG] historical_forecasts always retrains model

[evanwrm]

Describe the bug
historical_forecasts always retrains the model ignoring the retrain argument.

To Reproduce
This was likely introduced in #1465. Adding model = self.untrained_model() no longer sets self._fit_called. When evaluating whether to retrain not self._fit_called is always true.

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

from darts import TimeSeries
from darts import models
from darts.datasets import AirPassengersDataset

series_air = AirPassengersDataset().load().astype(np.float32)
model = models.LinearRegressionModel(lags=4)

model.historical_forecasts(
    series_air,
    start=0.9,
    retrain=False, # <-- this is ignored
)

Expected behavior
forecasting models to respect the retrain parameter.

System (please complete the following information):

• Python version: 3.9.16
• darts version 0.23.1

[dennisbader]

We introduced this in #1465 to avoid mutating the underlying model object.
The model in this case gets retrained because it was not fit before (Darts requires the models to be fit before prediction).
You are right that it is an unexpected behavior that if retrain=False, the model still gets retrained.

We should raise an error when retrain=False and the model was not fit before.

You can avoid the retraining by fitting the model once before calling historical_forecasts()