Describe the feature or idea you want to propose

Introduce a new base class, BaseDeepForecaster, to the aeon.forecasting.deep_learning module. This class provides a clean foundation for implementing deep learning-based forecasting models using TensorFlow/Keras.

Describe your proposed solution

Currently, aeon lacks a unified base class tailored for deep learning forecasting pipelines.

The BaseDeepForecaster is built on top of BaseForecaster while supporting key DL workflows. It can act as a foundation for models like Informer, TCN and DeepAR forcasters.

The proposed BaseDeepForecaster would:

• Inherit from BaseForecaster, passing horizon and axis via constructor.
• Define the abstract method _build_model(input_shape) for subclasses to implement specific architectures.
• Handle input conversion, sequence generation, model compilation, and training in _fit.
• Use the last observed window for forecasting in _predict.
• Support configuration via parameters: horizon, window, batch_size, epochs, verbose, optimizer, loss, and random_state.

Describe alternatives you've considered, if relevant

No response

Additional context

No response