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.).
wasm32 support
On wasm32-* targets the underlying download stack (hf-hub, dirs,
tokio::fs) is not available, so [ModelCache] is a stub that always
returns [CacheError::Unsupported]. Browser/Worker callers should obtain
model bytes through a different mechanism (e.g. fetch() on the JS side,
pre-bundled assets, or a manually populated cache directory).