feat: add Ollama GGUF runtime support

[FenjuFu]

Summary

• add a dependency-free GGUF v2/v3 metadata parser that never loads tensor weights and recognizes Q4_0, Q4_K_M, Q5_K_M, Q6_K, and Q8_0
• add local/remote Ollama REST support for server discovery, model listing, progress-aware pulls, loading, and deterministic generation
• add a CLI for GGUF inspection/discovery/import and Ollama-managed models
• keep local GGUF, Ollama-managed storage, and existing Safetensors training checkpoints explicitly separate
• document the workflow in English and Chinese and add a Python 3.10-3.12 CI matrix

Repository scope

Domux currently has no application UI or model-management service. This PR provides a reusable CLI/library entry point instead of introducing an unrelated UI framework.

Validation

• python -m unittest discover -s tests -v — 6 passed
• python -m compileall -q domux_runtime tests
• Ruff check and format — passed
• git diff --check — passed

Partially addresses #8 (criteria 1, 3, 5, 6). The UI upload path (criterion 2) and the in-place weight-loader memory optimization (criterion 4) are intentionally out of scope here — see the scope note above and the follow-up comment on #8.