Shared model download and cache layer for Blazen local-inference backends.
Provides [ModelCache] for downloading and caching ML models from
HuggingFace Hub. Designed to be shared by all
local-inference backends (fastembed, mistral.rs, whisper.cpp, etc.).