feat: support rollout importance sampling modified from verl

[RangiLyu]

Adapt VeRL's rollout importance sampling method

Modify from https://github.com/volcengine/verl/blob/main/verl/trainer/ppo/mismatch_helper.py

Main differences:

• Adapt to packed sequences in Xtuner
• Support TIS and token mask at the same time with different thresholds. In this way, we can easily implement the IcePop trick from Ring

Usage

1. TIS

loss_cfg = GRPOLossConfig(
    policy_loss_cfg=dict(
        cliprange_high=0.28,
        cliprange_low=0.2,
        loss_type="vanilla",
        clip_ratio_c=10.0,
        log_prob_diff_min=-20.0,
        log_prob_diff_max=20.0,
    ),
    ignore_idx=-100,
    use_kl_loss=False,
    kl_loss_coef=0.0,
    kl_loss_type="low_var_kl",
    mode="chunk",
    chunk_size=512,
    rollout_is=RolloutImportanceSampling(
        rollout_is_level="token",
        rollout_is_mode="truncate",
        rollout_is_threshold=(2, 0.5),
    ),
)

2. IcePop

loss_cfg = GRPOLossConfig(
    policy_loss_cfg=dict(
        cliprange_high=0.28,
        cliprange_low=0.2,
        loss_type="vanilla",
        clip_ratio_c=10.0,
        log_prob_diff_min=-20.0,
        log_prob_diff_max=20.0,
    ),
    ignore_idx=-100,
    use_kl_loss=False,
    kl_loss_coef=0.0,
    kl_loss_type="low_var_kl",
    mode="chunk",
    chunk_size=512,
    rollout_is=RolloutImportanceSampling(
        rollout_is_level="token",
        rollout_is_mode="both",
        rollout_is_threshold=(5, 0.5),
        rollout_is_mask_threshold=(5, 0.5),
    ),
)