[forecast] Allow specifying individual horizons

[christian-adam]

Description

Currently, for a multi-step horizon, MLForecast allows the recursive and the direct strategy. As the models for the direct strategy are independent of each other, training a model for every step up to max_horizon can create a lot of overhead when only some of those models are actually required. Say I want to predict 7 and 14 days in advance - the direct strategy would train 14 models, but I would only be needing two of those.

I propose a horizons parameter to be added to MLForecast and LightGBMCV that allows specifying directly which steps to train on, like so:

fcst = MLForecast(
    models=lgb.LGBMRegressor(random_state=0, verbosity=-1),
    freq=1,
    num_threads=1,
    # max_horizon=14,
    horizons=[7,14]
)

Use case

No response