feat: add GLM-4.7-Flash (glm4_moe_lite) model support

[Realmhang]

PR Category

Core

PR Type

New Features

Description

• Add support for GLM-4.7-Flash (glm4_moe_lite) model, which combines MLA (Multi-head Latent Attention) from DeepSeek V2/V3 with MoE
  architecture from GLM-4 MoE
• Add config bridge Glm4MoeLiteConfig extending Glm4MoeConfig with MLA fields, DSA Indexer fields, and MTP support
• Add model implementation (Glm4MoeLiteForCausalLM) inheriting from both glm4_moe and deepseek_v2 components
• Register model and config in plugin entry point

Changes

[CLAassistant]

All committers have signed the CLA.