feat: support GLM-5-FP8 in HYBRID mode

[jasonqinzhou]

All GLM-5-specific MLA params are correctly read. Here's a summary of what was done:

New files:

• model_configs/zai-org--GLM-5-FP8_config.json — 78 layers, 64 heads, 256 experts, MLA with q_lora_rank=2048,
  qk_nope=192, v_head=256

common.py:

• Added DeepSeekMLAConfig dataclass (5 MLA params)
• Registered GlmMoeDsaForCausalLM → "DEEPSEEK" family
• Added zai-org/GLM-5-FP8 to DefaultHFModels

utils.py:

• For any DEEPSEEK-family architecture with kv_lora_rank in config, populate extra_params as DeepSeekMLAConfig

models.py:

• All 3 DEEPSEEK model classes (DeepSeekModel, TrtllmWideEPDeepSeekModel, WideEPDeepSeekModel) now accept mla_config as
  a parameter
• Replaced all 7 hardcoded MLA constants in DeepSeekModel and TrtllmWideEPDeepSeekModel with expressions derived from
  mla_config (defaulting to DeepSeek-V3 values for backward compat)

  ┌─────────────────────────┬───────────────────────────────┬────────────────────────────────────┐                       
  │          Model          │         Architecture          │             Supported?             │
  ├─────────────────────────┼───────────────────────────────┼────────────────────────────────────┤                       
  │ GLM-5-FP8               │ GlmMoeDsaForCausalLM          │ ✅ Yes                             │                       
  ├─────────────────────────┼───────────────────────────────┼────────────────────────────────────┤
  │ GLM-4.7-Flash           │ Glm4MoeLiteForCausalLM        │ ✅ Yes (just added)                │
  ├─────────────────────────┼───────────────────────────────┼────────────────────────────────────┤
  │ GLM-4.7-FP8             │ Glm4MoeForCausalLM            │ ✅ Yes (just added)                │
  ├─────────────────────────┼───────────────────────────────┼────────────────────────────────────┤
  │ GLM-4.6V-Flash-MLX-8bit │ Glm4vForConditionalGeneration │ ❌ No (vision + MLX/Apple Silicon) │
  ├─────────────────────────┼───────────────────────────────┼────────────────────────────────────┤
  │ GLM-4.6V-Flash-MLX-6bit │ Glm4vForConditionalGeneration │ ❌ No (vision + MLX/Apple Silicon) │
  └─────────────────────────┴───────────────────────────────┴────────────────────────────────────┘ 

[tianhaox]

glm 5 needs DSA, same as DS V3.2

[tianhaox]

suggest we remove the wideep support in this pr. Do more complete design in a seperate PR.

[tianhaox]

This is incorrect.
https://huggingface.co/zai-org/GLM-5-FP8/blob/main/config.json

[tianhaox]

i will suggest we do manual copy paste to avoid illusion.

[tianhaox]

this is incorrect. without quant field.

[tianhaox]

I think this whole PR needs a better design of DEEPSEEK_V32 model family. DS V32 and GLM-5 will share the model family. It uses DSA+MoE. I suggest we cancel this PR. Redesign after the DSA Op PR I've created. I will start a PR manually to better support this DEEPSEEK_V32 model family