[Fix] Run GLM-5.2 shared MTP for 7 logical steps

[Knlife]

Motivation

The official GLM-5.2 release uses shared recurrent MTP steps. The official blog[1] reports 7 MTP steps for both training and inference, while the released checkpoint retains num_nextn_predict_layers=1. XTuner's Glm52MoEConfig.from_hf() currently maps this field directly to MTPConfig.num_layers, so the canonical SFT recipe constructs 1 physical shared MTP layer but executes it for only one logical step.

[1] https://huggingface.co/blog/zai-org/glm-52-blog
[2] https://huggingface.co/zai-org/GLM-5.2/raw/main/config.json

Changes

Override examples/v1/config/sft_glm5p2.py with

model_cfg.mtp_config = MTPConfig(
    num_layers=7,
    share_weights=True,
    detach_mtp_lm_head_weight=False,
    detach_mtp_inputs=False,
    loss_scaling_factor=0.1,
)

Preserve num_nextn_predict_layers=1 so checkpoint construction and strict loading continue to use 1 physical MTP layer. Add a regression test for the recipe configuration. The existing shared-weight implementation in MTPBlock and GLM-5.2 model construction remains unchanged.

Scope and compatibility

This change is limited to the canonical GLM-5.2 SFT recipe. Generic Hugging Face conversion, existing CE loss behavior, MTP loss scaling, and strict_load=True behavior are unchanged. Other autotest and RL configurations, including explicit MTP4 cases, are intentionally left unchanged. This PR does not implement or modify the end-to-end TV loss objective.

Validation

python -m py_compile and git diff --check passed. A dependency-free recipe smoke test passed. Full pytest and CUDA/strict-load integration tests require the project runtime dependencies and checkpoint, so they are left to CI.