Preserve GLM-5.2 Megatron accuracy-compatible specs

[zrr1999]

Summary

• infer MTP layer count from the GLM model configuration
• prefer the configured routed-expert count during Megatron construction
• preserve accuracy-compatible DSA/RMSNorm submodule specs when mcore-bridge rewrites the GLM DSA spec

These changes let the existing ms-swift + mcore-bridge path construct the reduced official GLM-5.2 profile without replacing Megatron modules with toy/eager stand-ins.

Validation

• official-weight GLM-5.2 checkpoint conversion/load completed through the bridge
• the resulting Megatron-Core model ran CUDA/bfloat16 forward/backward on NVIDIA H800
• forward loss matches the independent Paddle path at 12.416634559631348

Status

Draft while strict backward, optimizer, and 100-step parity remain under validation.