[Feat] Add Sampled-token and top-k OPD Support to RL Training - #1977

  loss_cfg = DistillationLossConfig(
      policy_loss_cfg=policy_loss_cfg,
      loss_mode="k1",  # k1, forward, reverse, forward_kl_topk
      use_policy_gradient=True,
      task_adv_weight=0.0,
      distillation_loss_weight=1.0,
      top_k=64,
      log_prob_min_clamp=-10.0,
      loss_max_clamp=10.0,
  )

distillation_config = DistillationConfig(
    loss_config=loss_cfg,
    teachers=[
        RolloutTeacherConfig(
            name="teacher",
            launch_config=RolloutTeacherLaunchConfig(
                model_path=teacher_model_path,
                server_port=13141,
            ),
        ),
    ],
    data_source_teacher_map={
        "math": "teacher",
    },
)

distillation_config = DistillationConfig(
    loss_config=loss_cfg,
    teachers=[
        TrainTeacherConfig(
            name="teacher",
            model_path=teacher_model_path,
            model_cfg=teacher_model_cfg,
            fsdp_cfg=teacher_fsdp_cfg,
        ),
    ],
    data_source_teacher_map={
        "math": "teacher",
    },
)