Add Darts Torch Forecasting dataset to save and load models

[SzymonCogiel]

Description

I'm finding it challenging to effectively save and load Darts Torch Forecasting Models within a Kedro pipeline using kedro-datasets. Current dataset options, like PickleDataset and PyTorchDataset, lack compatibility with the specific requirements and architecture of Darts models especially those utilizing PyTorch as a backend. Darts Torch Forecasting Models, including RNN, TCN, and Transformer models and more, incorporate essential features such as covariate dependencies, automatic checkpointing, and probabilistic forecasting, which are vital for accurate model persistence and reloading.

Context

Kedro is widely adopted for managing data pipelines, and adding native support for Darts Torch models would enable seamless integration of advanced time series forecasting capabilities. This addition would benefit users by streamlining model persistence and improving compatibility for Darts-specific features within Kedro pipelines.

Possible Implementation

A potential implementation could involve creating a DartsTorchDataset class within Kedro Datasets Experimental that leverages Darts' native save() and load() methods. This class should also support checkpointed loading through Darts’ load_from_checkpoint() where applicable. It would handle the specific requirements for Darts models, including covariate management and probabilistic forecasting support, ensuring compatibility across a wide range of Darts Torch Forecasting Models.