Tabular forecasting models would occasionally fail when a non-boolean known real covariate was mistakenly interpreted as a boolean covariate.

MWE:

import pandas as pd
from autogluon.timeseries import TimeSeriesPredictor
N = 30
df = pd.DataFrame(
    {
        "item_id": ["A"] * N,
        "timestamp": pd.date_range("2020-01-01", freq="D", periods=N),
        "target": np.random.normal(size=N),
        "feat": np.tile([5, 0, 0], int(N/3)),
    }
)
predictor = TimeSeriesPredictor(known_covariates_names=["feat"]).fit(df, hyperparameters={"RecursiveTabular": {}})

This code will fail with the following exception

	Warning: Exception caused RecursiveTabular to fail during training... Skipping this model.
	"['__scaled_feat'] not in index"

This happens because during fit(), the feature feat is interpreted as non-boolean, so a scaled copy of the feature is added. At predict time, when transforming known_covariates (containing one row with only the 0 value), the feature was interpreted as boolean, so no scaled version of the feature was added. This results in the model failing at prediction time.

Description of changes:

• This PR fixes the above bug.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.