test: add test for per-attack image directories and improve error handling

- Add test_generate_test_cases.py to verify each attack uses its own image directory
- Improve error messages in unified_registry for failed imports with more context
- Add null checks in generate_test_cases.py for attack initialization failures
- Reduce excessive max_tokens in judge_utils from 130000 to 10000 for efficiency
- Add missing ThreadPoolExecutor import in qr/attack.py
- Fix missing newline in pyproject.toml

evaluators/judge_utils.py

@@ -491,15 +491,15 @@ Your judgment must be more than a simple formula. You are to weigh the evidence
 
     # Generate evaluation responses - use adapter function
     evaluation_harmfulness_response = _call_model_generate(
-        judge_client, evaluate_harmfulness_messages, max_tokens=130000, temperature=0
+        judge_client, evaluate_harmfulness_messages, max_tokens=10000, temperature=0
     )
 
     evaluation_alignment_response = _call_model_generate(
-        judge_client, evaluate_alignment_messages, max_tokens=130000, temperature=0
+        judge_client, evaluate_alignment_messages, max_tokens=10000, temperature=0
     )
 
     evaluation_detail_response = _call_model_generate(
-        judge_client, evaluate_detail_messages, max_tokens=130000, temperature=0
+        judge_client, evaluate_detail_messages, max_tokens=10000, temperature=0
     )
 
     # Extract scores and reasons
@@ -527,7 +527,7 @@ Your judgment must be more than a simple formula. You are to weigh the evidence
     ]
 
     final_evaluation_response = _call_model_generate(
-        judge_client, final_message, max_tokens=130000, temperature=0
+        judge_client, final_message, max_tokens=10000, temperature=0
     )
     final_score, final_reason = extract_evaluation_score(final_evaluation_response)