bug (tft) - wrong num_encoder_steps for retail dataset favorita

I guess we are taking 90 days history to forecast for next 30 days. So shouldn't the num_encoder_steps be 90 instead ?
It seems for other datasets num_encoder_steps are defined correctly, but wrong for retail dataset (favorita)

tft/data_formatters/favorita.py

# Default params
def get_fixed_params(self):
"""Returns fixed model parameters for experiments."""

fixed_params = {
'total_time_steps': 120,
'num_encoder_steps': 30,
'num_epochs': 100,
'early_stopping_patience': 5,
'multiprocessing_workers': 5
}

return fixed_params

Start in tft/data_formatters/favorita.py at get_fixed_params and compare num_encoder_steps with the stated 90-day history and 30-day forecast, as well as the settings for other datasets. Confirm the intended split before changing the value; done means the Favorita configuration uses the correct encoder length for its forecast horizon.

Written by the indexing model from the issue text.