support mtp layer support for qwen3.5 series models

[zpltys]

support qwen3.5 models' mtp layer.
I have test the correctness in example/qwen3_5/test_mtp_logits.py file and e2e sft training on qwen3.5 35ba3b and 9b

[ArronHZG]

How does the mtp_loss_scaling_factor take effect?
Also, how do I load a model and disable MTP?

[ArronHZG]

def _build_config(self):
    """Override to add MTP configuration."""
    hf_config = self.hf_config

    # Add MTP configuration if present
    mtp_args = {}
    if "num_nextn_predict_layers" in hf_config:
        mtp_args["mtp_num_layers"] = hf_config.num_nextn_predict_layers
        mtp_args["mtp_loss_scaling_factor"] = self.extra_args.get("mtp_loss_scaling_factor", 0.1)

    return self._build_base_config(
        add_qkv_bias=True,
        qk_layernorm=False,
        **mtp_args,
    )

just like this.

[zpltys]

How does the mtp_loss_scaling_factor take effect? Also, how do I load a model and disable MTP?

I have fix this