[TD] Decouple AiExecutor Batch Pattern

[ArthurCRodrigues]

Summary

AI-enabled tests currently register themselves with AiExecutor, return placeholder results, and rely on GraderService to detect executors after grading to send a batch request. Results are then mutated in place, breaking the step-by-step pipeline model.

Scope

• Redesign AI execution so batch processing happens in a dedicated pipeline-aware component (e.g., separate step or executor strategy) instead of hidden inside test functions.
• Remove the hasattr(test_func, "executor") check in GraderService and disallow in-place mutation of TestResult objects.
• Ensure AI tests follow the same contract as other tests (execute → return result).

Impacted Files

• autograder/utils/executors/ai_executor.py
• autograder/services/grader_service.py
• Templates/tests using AI executors

Risks

• AI grading may regress if batch semantics change; add targeted integration tests.

Acceptance Criteria

• AI execution no longer hijacks test functions; results flow through the pipeline like other tests.
• GraderService has no executor-specific checks.

References

• docs/roadmaps/TECHNICAL_DEBT_ROADMAP.md (Item 21)