Review models architecture

[ben-jy]

According to potential models that will be benchmarked, and the way we will benchmark them, architecture should be reviewed. We must define if it makes sense to create a new interface for benchmarked models, or if we use interface/class already defined in ontime (AbstractModel/Model, see https://github.com/ontime-re/ontime/tree/develop/src/ontime/core/modelling)

For the benchmark models, we need :

• an evaluation method, to be performed in rolling window mode
• a basic fit method, with training managed by the trained model
• a predict method, which can predict on a new series as well as on the one the model has been trained on
• a flag to determine the mode of evaluation (zero-shot, full-shot, few-shot)
• the possibility to load a checkpoint, particularly for foundation models
• univariate model should be able to perform multi-univariate predictions when given input series is multivariate
• a means of resetting or reinstating the model between runs (Improve benchmark objects instanciation #68)

This issue in linked to issue #46

[ben-jy]

For now, we created an interface AbstractBenchmarkModel (see https://github.com/ontime-re/ontime/tree/48-benchmarking-fix-zero-few-shot/src/ontime/module/benchmarking) that must implement every benchmark models. This class is completely independent from existing ontime model.
We must therefore define if we keep the code this way or if we reuse existing class.

[ben-jy]

we do not need to specifically implement the possibility of loading of checkpoint, as we can let the specific wrapper to handle this with e.g. a "checkpoint" parameter.