feat: add sharding support for mlx-lm models

[andthattoo]

Integrate all MLX-LM model architectures with proper sharding augmentations for distributed inference in dnet.

Priority based on production deployments, HuggingFace downloads, and benchmark performance

• gpt_oss
• deepseek_v2
• deepseek_v3
• llama
• llama4
• qwen3
• qwen3_moe
• qwen3_next
• qwen2
• qwen2_moe
• internlm3
• gemma3
• gemma3_text
• gemma3n
• glm4
• glm4_moe
• olmo2
• olmo3

[GandalfTea]

We both added llama3 (#11 , #15 ) so we can resolve it. I have llama4, qwen3_moe (with dense experts), glm, glm4 and gpt_oss from the old PR, just need to adapt them to the new signature and add the unload function. Should I open a PR for this or do you want to do it in #15 ?

[andthattoo]

There were some issues with layer shrinking and quantization so I assume we add all the remaining models after #15 is finished.