feat: ModelAutoLoader node with auto-download from HuggingFace/ModelScope

- New QuantFunc Model Auto Loader node: select model series, transformer,
  prequant weights, and precision config from dropdowns
- Auto-detect GPU variant (50x-below/50x-above) based on SM version
- Auto-download base model, transformer, prequant, and precision-config
  from ModelScope or HuggingFace with resume support
- Resource cache: fetches available file listings from ModelScope on
  startup (background), cached to bin/<platform>/resource_cache.json
- Download completeness tracked via .quantfunc_download_complete marker
- Device dropdown shows detected GPU names instead of raw INT
- Data source dropdown (modelscope default, huggingface alternative)