Feat: sktime support

[spolisar]

sktime support

Added an adapter in timecopilot/models that can be used to work with sktime models using TimeCopilot.

Adding an SKTime forecaster currently requires the adapter be wrapped around a BaseForecaster (the abstract class sktime forecasters inherit from) by the user before passing it in a model list to a TimeCopilot agent. Setup for sktime models is currently being left to users as those arguments may vary according to the model and data. BaseForecaster can be added as a type hint if sktime is added to the requirements.

adapted_model = SKTimeAdapter(
    model=TrendForecaster() ,
    alias="TrendForecaster",
)

tc = timecopilot.TimeCopilot(
    llm="openai:gpt-4o",
    forecasters=[
        adapted_model,
    ],
)

Alias currently defaults to "SKTimeAdapter", should it be a required argument?

There currently is not support for levels and quantiles via the sktime adapter. It doesn't look like those are available for every sktime forecaster, so a decision on how to handle those cases should be made before supporting those features.

Exogenous data could be supported, and would also be a next step to take for sktime support. Exogenous data may have similar constraints to quantiles/levels.

Where should the sktime notebook be placed in the list of example notebooks?

[spolisar]

This should be moved over to a full PR once the notebook finds its place in the examples list.

Are more substantial doc strings also necessary?

Quantile support can be added after deciding how to handle the mixed support for that within sktime forecasters.

Exogenous data can also be a future feature, it may have similar support constraints to quantiles.