fix(reward): follow the GRPO reward contract in MathORM - #10266

What is changed / 改动内容

MathORM.__call__ now follows the reward-function contract that the GRPO dispatch uses, like every other
built-in ORM: it receives the completion texts and the dataset solution column, and still unwraps
InferRequest objects for programmatic callers (the same isinstance(..., str) branch ReactORM uses).

Why / 原因

--reward_funcs math crashed as soon as rewards were computed, because MathORM was the only registered ORM
left assuming List[InferRequest] while swift/rl_core/grpo_algorithm.py::compute_rewards_per_func passes
completions: List[str] positionally plus the dataset columns as kwargs:

• solution column (what the math datasets and the reward function guide document) → TypeError: MathORM.__call__() missing 1 required positional argument: 'ground_truths'
• a ground_truths column → AttributeError: 'str' object has no attribute 'messages' at swift/rewards/orm.py:444

Closes #10265

Test / 测试

• New TestMathORMCompletions cases in tests/utils/test_rewards.py (completion strings, solution kwarg,
  InferRequest): the module runs FAILED (errors=2) on dbc395ff without the fix and OK with it.
• End to end through the real dispatch (compute_rewards_per_func) with a solution column: crash before the
  change, [[1.0], [0.0]] after it for \boxed{3} vs 4 against \boxed{3}.
• tests/utils/test_math_orm_expressions.py, which calls MathORM with InferRequest objects, keeps passing.
• flake8, yapf and isort clean on both files.