refactor: generalize dataset indexing from language-based to dataset_id-based

[federetyk]

Addresses #33

Description

This PR generalizes dataset indexing within tasks from Language enum to arbitrary string identifiers (dataset_id). The current architecture limits each task to at most one dataset per language, which prevents supporting tasks with multiple monolingual datasets per language, cross-lingual datasets, or multilingual datasets.

The refactor introduces a languages_to_dataset_ids() method with a default 1:1 mapping that preserves backward compatibility for existing tasks. Tasks that require more complex dataset structures can override this method to return custom identifiers. A new get_dataset_language() method maps datasets back to their language for proper per-language result aggregation, returning None for cross-lingual or multilingual datasets.

Changes:

• Rename lang_datasets: dict[Language, Dataset] to datasets: dict[str, Dataset] in Task base class
• Add languages_to_dataset_ids(languages) -> list[str] method with default backward-compatible mapping
• Rename load_monolingual_data(language, split) to load_dataset(dataset_id, split) across all tasks
• Add get_dataset_language(dataset_id) -> Language | None method for per-language aggregation
• Add language field to MetricsResult to track dataset language
• Update _aggregate_per_language() to group by the language field, skipping datasets marked as cross-lingual or multilingual
• Update all task implementations to use the new method signature
• Add unit test for multi-dataset task scenarios
• Fix minor issues in some files in examples/

All tests pass, and the output of examples/run_multiple_models.py produces results consistent with the main branch.

Checklist

• Added new tests for new functionality
• Tested locally with example tasks
• Code follows project style guidelines
• Documentation updated
• No new warnings introduced