grade_answer_verl fails when rm_type="boxed_math" strips the \boxed{} wrapper upstream

The “boxed” reward shortcuts (rm_type="boxed_math", "boxed_dapo", etc.) already strip the LaTeX box around the response before calling grade_answer_verl. However, grade_answer_verl still assumes the solution string contains a \boxed{} and unconditionally calls extract_answer(solution_str). Because the wrapper has already been removed, extract_answer returns None, the helper immediately returns False, and the rollout get scored as incorrect even when the numeric answer matches the label.

Where

• slime/slime/rollout/rm_hub/init.py lines 28–37: the rm_type.startswith("boxed_") branch reassigns response = extract_boxed_answer(response) or "" before dispatching to grade_answer_verl.
• slime/slime/rollout/rm_hub/math_utils.py lines 487–494: grade_answer_verl always parses the response with extract_answer(solution_str).

Steps to Reproduce

1. Use rm_type="boxed_math" when calling async_rm.
2. Feed a sample whose response ends with Answer: \boxed{5}.
3. Because of the boxed prefix, init.py strips the wrapper, so solution_str is just "5" when it reaches grade_answer_verl.
4. extract_answer("5") returns None; the function returns False, yielding reward 0.

Start in slime/slime/rollout/rm_hub/init.py at the rm_type.startswith("boxed_") branch, then read grade_answer_verl in slime/slime/rollout/rm_hub/math_utils.py. Reproduce the boxed_math path with a response ending in Answer: \boxed{5}, and verify that the stripped response is graded correctly instead of returning False.