refactor(trainers): share auxiliary-role lifecycle

[yuetongz417]

Summary

This PR extracts the auxiliary model-role lifecycle shared by DPO and PPO into reusable role-aware trainer classes.

The new RoleAwareTrainerMixin centralizes:

• backend initialization and cleanup;
• model-role initialization through ensure_roles;
• role-based log-probability scoring;
• validation that returned log-probabilities align with the input token rows.

RoleAwarePolicyTrainer combines this shared role handling with PolicyOnlyTrainer, allowing rollout-based algorithms to reuse the same infrastructure.

DPO now inherits the generic role-aware mixin for its frozen reference model, while PPO inherits the role-aware policy trainer for its actor, reference, critic, and reward roles. Their existing training behavior is unchanged.

This refactor also provides a reusable foundation for future teacher-scored workflows such as OPD, where a trainable student performs rollouts and a frozen teacher scores the generated tokens.

Testing

Added CPU tests covering:

• DPO and PPO inheritance from the shared role-aware classes;
• role initialization order;
• backend cleanup when role initialization fails;
• forwarding of scoring microbatch configuration;
• validation of missing or misaligned log-probability rows.