refactor(llm): Split model.py by model architecture

[m96-chan]

Problem

src/pygpukit/llm/model.py is 1501 lines with multiple model implementations mixed.

Current State

src/pygpukit/llm/
└── model.py  (1501 lines - all models)

Proposed Structure

src/pygpukit/llm/models/
├── __init__.py          (exports)
├── base.py              (BaseModel, common interfaces)
├── qwen.py              (QwenModel, Qwen2, Qwen3)
├── llama.py             (LlamaModel)
├── moe.py               (MoE model support)
└── registry.py          (model registry, auto-detection)

Benefits

• Each model in its own file
• Easier to add new model architectures
• Clear inheritance hierarchy
• Model-specific optimizations isolated

Related

• refactor(llm): Split model.py by model architecture #141 (layers.py split)

[m96-chan]

Completed in PR #168