feat: optimize the groupmatmul operator.

[LMX-xin]

• Change 1: Modify the down weight in deepseek_v2_decoder_layer.cpp to NZ format.
• Change 2: In the ATB plugin, activate the shape based on the input to select different operators.
• Performance:

When using groupmatmul with dp=1 and ep=1, tpop decreased by 5%, ttft decreased by 5%, and throughput improved by 5%.

[read-the-docs-community]

Documentation build overview

📚 xllm_pub | 🛠️ Build #29273701 | 📁 Comparing 82d1feb against latest (43befa4)

🔍 Preview build

Show files changed (3 files in total): 📝 2 modified | ➕ 0 added | ➖ 1 deleted

File	Status
404.html	📝 modified
index.html	📝 modified
zh/features/moe_params/index.html	➖ deleted

[yq33victor]

LGTM