[Auto] Auto* models can not be used with any config as a subset of default_config

[yarnabrina]

Description

Few models (e.g. TFT) takes required positional arguments (e.g. input_size). If user passes a config to AutoTFT that does not have this key, training fails with following as the root cause:

TypeError: TFT.init() missing 1 required positional argument: 'input_size'

Here's a reproducible example:

import pandas
from neuralforecast import NeuralForecast
from neuralforecast.auto import AutoTFT
from ray import tune

from sktime.datasets import load_longley
from sktime.split import temporal_train_test_split

y, X = load_longley()
y_train, y_test, X_train, X_test = temporal_train_test_split(y, X, test_size=4)

algorithm = AutoTFT(
    4,
    config={"max_steps": tune.choice([5, 10, 15]), "random_seed": tune.choice([0])},
    backend="ray",
)
model = NeuralForecast([algorithm], "A-DEC")

train_data = {
    "unique_id": 1,
    "ds": y_train.index.to_timestamp(freq="A-DEC").to_numpy(),
    "y": y_train.to_numpy(),
}
for column in X.columns:
    train_data[column] = X_train[column].to_numpy()

train_dataset = pandas.DataFrame(data=train_data)

model.fit(df=train_dataset)
# RuntimeError: No best trial found for the given metric: loss. This means that no trial has reported this metric, or all values reported for this metric are NaN. To not ignore NaN values, you can set the `filter_nan_and_inf` arg to False.

It will be helpful to document mandatory keys to be passed in config in the documentation of Auto* models.

If it is possible to allow users to pass few arugments to the underlying model directly without tuning, for __init__ (where tuning is not necessary) or to fit or predict (for example configurations for PyTorch Lighning's trainer), it will be very useful.

Link

No response