Fix prediction target encoded

[Irozuku]

This pull request refactors how predictions are processed across different machine learning tasks in the DashAI backend. It introduces a unified process_predictions method in the base task class, implements task-specific logic for classification tasks, and updates related endpoints and job logic to use this new structure. The changes improve code modularity, make prediction handling more consistent, and simplify the addition of new task types.

Task Abstraction and Prediction Processing

• Added an abstract process_predictions method to BaseTask, requiring all tasks to define how predictions are processed.
• Implemented process_predictions in ClassificationTask, converting model outputs to class labels using dataset metadata, and updated all classification task types (ImageClassificationTask, TabularClassificationTask, TextClassificationTask) to inherit from ClassificationTask instead of BaseTask. [1] [2] [3] [4]
• Implemented process_predictions for regression and translation tasks to simply return predictions as-is, maintaining compatibility. [1] [2]

Dataset Utilities and Label Encoding

• Added the encode_labels utility to encode categorical columns into numerical labels and generate a ClassLabel feature, used by classification tasks for prediction processing. [1] [2]

Prediction Job and API Endpoint Updates

• Updated the prediction job logic to use the new process_predictions method, loading the appropriate task from the registry and handling dataset loading errors more robustly. [1] [2]
• Refactored the /predict_summary API endpoint to extract prediction metadata, determine output data types using the task class, and compute class distributions for classification tasks using the updated structure. [1] [2]

Before

After