feat:enable v2 training pipeline with controller parity

[sitabulaixizawaluduo]

Changes

• GatewayTrainController: add version management (set_version / get_version), connect_engine, clear_batches; persist guard addresses for later port allocation; unify HTTP client session (follow-up to feat: controller v2 refactor #1354).
• Rollout entrypoints (rl_trainer, sglang_remote, vllm_remote): route to RolloutControllerV2 when config._version == "v2".
• Weight update controller: small adjustments to the connect method for the v2 path.
• Examples: add an agent config section to all examples/math/*.yaml; switch the default workflow in gsm8k_rl.py to MathAgent.
• Reward: remove obsolete get_custom_reward_fn from areal/reward/__init__.py.

[gemini-code-assist]

Code Review

This pull request introduces 'v2' controller support, integrating RolloutControllerV2 and implementing RL parity methods like connect_engine and update_weights in the GatewayTrainController. It also refactors weight update logic, removes legacy reward utilities, and updates example configurations and tests. Feedback identifies a critical AttributeError in connect_engine due to an undefined attribute and notes that update_weights incorrectly invokes asynchronous methods synchronously while containing redundant generation control logic that may conflict with the trainer's state.

[gemini-code-assist]

The attribute inference_worker_urls is not defined in RolloutControllerV2, which will cause an AttributeError at runtime. You should use the internal _inf_addrs attribute or add a public property to RolloutControllerV2 to expose these URLs.

Suggested change

	inference_urls: list[str] = rollout.inference_worker_urls
	inference_urls: list[str] = rollout._inf_addrs

[gemini-code-assist]

This block has multiple issues:

1. The calls to pause_generation() and continue_generation() are asynchronous but called synchronously, returning coroutines without executing them.
2. These calls are redundant and potentially harmful here. PPOTrainer.train already manages the rollout pause/resume state. In RolloutControllerV2, pause() correctly stops generation. Resuming it here via continue_generation() would break the trainer's expectation that inference remains paused during the subsequent save and evaluation steps.
3. The assert should be replaced with a proper runtime check as assertions can be disabled in production.

        if meta.version is None or meta.version <= 0:
            raise ValueError(f"meta.version must be a positive integer, got {meta.version}")
        result = self._weight_update_ctrl.update_weights(version=meta.version)

[TaoZex]

LGTM

[fishcrap]

LGTM