Combine every eval ever

[Chibukach]

This PR implements evaluation result aggregation functionality to average results across multiple random seeds. When running evaluations with different random seeds (for statistical reliability), this feature automatically combines the results.

Changes include:

• Core functionality to read and group evaluation results by benchmark
• Logic to compute average accuracy scores across different random seeds
• Support for both simple tasks and tasks with subtasks
• Updated CLI interface
• Tests for the seed merging functionality