feat: separate task execution from evaluation in Experiment with run_tasks/run_evaluators

[afarntrog]

Description

Refactor the Experiment class to separate task execution from evaluation into independent, composable steps, and fix async result ordering.

• Add run_tasks() and run_tasks_async() methods that execute tasks against cases and return EvaluationData without running evaluators
• Add run_evaluators() and run_evaluators_async() methods that evaluate pre-computed EvaluationData without executing tasks, enabling reuse of task results across multiple evaluation runs
• Rewrite run_evaluations() and run_evaluations_async() as convenience wrappers that compose the two steps
• Preserve result ordering in async workers by using index-based placement into a pre-allocated list instead of appending to a shared list
• Switch from asyncio.iscoroutinefunction() to inspect.iscoroutinefunction() for more reliable coroutine detection
• Extract shared logic into helper methods (_build_reports, _init_evaluator_data, _log_evaluation_to_cloudwatch, _set_task_span_attributes, _extract_task_error, _build_failed_evaluation_data)
• Failed task executions now produce EvaluationData with actual_output=None and the error in metadata["task_error"] instead of raising
• Always include session_id in evaluation metadata

Related Issues

#94

Documentation PR

Type of Change

New feature

Testing

Added 15 new tests in tests/strands_evals/test_experiment.py covering:

• run_tasks: basic output, dict output with trajectory/interactions, task failure handling
• run_evaluators: pre-computed data evaluation, evaluator error isolation, multiple evaluators
• Roundtrip: run_tasks → run_evaluators pipeline, serialization/deserialization roundtrip
• Async variants: run_tasks_async with sync and async tasks, failure handling, run_evaluators_async with pre-computed data and evaluator failures

Updated existing tests to reflect span naming changes (run_task/eval_case instead of execute_case) and the inclusion of session_id in metadata.

• I ran hatch run prepare

Checklist

• I have read the CONTRIBUTING document
• I have added any necessary tests that prove my fix is effective or my feature works
• I have updated the documentation accordingly
• I have added an appropriate example to the documentation to outline the feature, or no new docs are needed
• My changes generate no new warnings
• Any dependent changes have been merged and published

---

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[afarntrog]

Using the index instead of just appending is needed to maintain order. Previously, order was not guaranteed and you can have the following outcome:

3 cases and 2 workers:                                        
  
  Time    Worker-0                    Worker-1                                                                                             
  ────    ────────                    ────────                    
  t0      get_nowait() → case_0                                                                                                            
  t1      await task(case_0)...       get_nowait() → case_1                                                                                
  t2      (task_0 still running)      await task(case_1)...                                                                                
  t3      (task_0 still running)      task_1 completes!                                                                                    
  t4      (task_0 still running)      results.append(result_1)  ← index 0 in list!                                                         
  t5      task_0 completes!           get_nowait() → case_2                                                                                
  t6      results.append(result_0)    await task(case_2)...                                                                                
          ↑ index 1 in list!