refactor(llm): Simplify loader.py - extract format-specific loaders

[m96-chan]

Problem

src/pygpukit/llm/loader.py is 1015 lines handling multiple model formats.

Current State

src/pygpukit/llm/
└── loader.py  (1015 lines - all formats)

Proposed Structure

src/pygpukit/llm/loaders/
├── __init__.py          (unified load interface)
├── base.py              (BaseLoader)
├── safetensors.py       (SafeTensors loader)
├── gguf.py              (GGUF loader - future)
├── hf_hub.py            (HuggingFace Hub download)
└── weight_map.py        (weight name mapping)

Benefits

• Each format loader isolated
• Easier to add new formats (GGUF, AWQ, etc.)
• Clear loader interface
• Better error handling per format

API

from pygpukit.llm.loaders import load_model

# Auto-detect format
model = load_model('path/to/model')

# Explicit format
model = load_model('path/to/model', format='safetensors')

Related

• refactor(llm): Split model.py by model architecture #141 (model.py split)
• refactor(llm): Clean up __init__.py - move logic to dedicated modules #143 (init.py cleanup)

[m96-chan]

Completed in PR #168