feature: integrate tile db loading in data_loader

[VittorioRossi]

No description provided.

[giacomo-ciro]

I would modify the dataloader to return:

return {
            "input_ids": input_ids,             # [max_length,]: LongTensor
            "eos_labels": eos_labels,           # [max_length,]: LongTensor
            "response_length": response_length, # [1,]: FloatTensor
            "input_mask": input_mask            # [max_length,]: BoolTensor,
            ++ "input_embeddings: input_embeddings # [max_length, d_model]: FloatTensor
        }

Effectively, the input_ids won't be needed anymore, but I suggest we keep them just in case (the pipeline is already in place, just ignore them).

And also modify the models' forward pass (LLaDaRegressor.forward() and LLaDaClassifier.forward() in models/llada.py) accordingly to take directly the embeddings as input and work on those.

As of now, the model takes only input_ids and calls hidden_state = self.llada.get_last_hidden_state(input_ids), since we have the embeddings stored on disk we can skip this part and directly go with hidden_state = input_embeddings.

[giacomo-ciro]

@VittorioRossi ho fatto il wrapper al trainer in trainer.py e fatto lo script train.py che ora e' super clean (5 righe).

Rimane da adattare il forward pass dei modelli e testare tutto (io non ho ancora runnato il mio perche mancano delle cose,quindi ci saranno bug, ma una volta terminato tutto il refactor li fixiamo