refactor(glm): reuse DeepseekV3 base implementation via template methods

[wtsung]

Summary

Refactor GlmMoeDsaDecoderLayer and GlmMoeDsaModel to fully leverage inheritance from their DeepseekV3 base classes.

This change introduces two template methods in the DeepseekV3 implementation so that subclasses only need to customize the components that differ, eliminating approximately 170 lines of duplicated code without changing behavior.

Changes

deepseek_v3.py

Introduce two extension points for subclasses:

• Add DeepseekV3DecoderLayer._create_attn() to allow subclasses to customize the attention implementation.
• Add DeepseekV3Model._create_decoder_layer() to allow subclasses to customize decoder layer construction.

glm5.py

Simplify the GLM-specific subclasses by relying on the inherited implementation wherever possible.

GlmMoeDsaDecoderLayer

• Call super().__init__() instead of duplicating the parent initialization.
• Override only _create_attn() to instantiate GlmMoeDsaAttention.
• Remove duplicated implementations of forward(), forward_mlp(), and initialization logic, all of which are identical to DeepseekV3DecoderLayer.

GlmMoeDsaModel

• Override only _create_decoder_layer() to construct GlmMoeDsaDecoderLayer.
• Remove the duplicated __init__() implementation and inherit the base initialization directly.

Additional cleanup

• Inline single-use static helper methods.
• Initialize _retired_decode_workspaces in __init__().
• Remove the unused parameter from _resolve_decode_q_len().
• Rename _qkv_width to qkv_width.

Verification

The refactor is behavior-preserving:

• The removed forward() and forward_mlp() implementations in GlmMoeDsaDecoderLayer are identical to those inherited from DeepseekV3DecoderLayer.
• The removed GlmMoeDsaModel.__init__() performs the same initialization sequence as DeepseekV3Model.__init__().
• The only functional customization retained in the GLM subclasses is the creation of the attention module and decoder layer via the newly introduced template methods.