Add gated GLM-5.2 bit-exact compatibility paths

[zrr1999]

• I, the PR author, have personally reviewed every line of this PR.

What does this PR do?

Draft。只改 Megatron-LM。三个兼容 flag 和 native_unfused_adamw 默认 False，关闭时走原来的 fused/native 路径。

• --norm-accuracy-compatible：显式 fp32 RMSNorm。DSA qk-norm 与 MTP layer-norm spec 可选用。
• --dsa-accuracy-compatible：关掉 fused DSA kernel；unfused 路径用显式 masked softmax backward。
• --router-accuracy-compatible：router gating 用 fp32 torch.mm。
• native_unfused_adamw：torch.optim.AdamW(foreach=False, fused=False)。

MTP spec 把 config 传入 get_mtp_layer_spec*，否则 MTP 选不到上面的 RMSNorm。

测试：RMSNorm 兼容路径、DSA unfused softmax backward、router fp32 gating、MTP spec 使用传入的 config。

Issue tracking

Linked issue: 无

Contribution process

Pre-checks

• I have added relevant unit tests
• I have added relevant functional tests
• I have added proper typing to my code Typing guidelines
• I have added relevant documentation
• I have run the autoformatter.sh on my PR