clef_mlx.py: use the repo name (not the cache snapshot hash) as the default model name

clef_mlx.py: default --model works inside Hub cache snapshots (symlinked files)

clef_mlx.py: add 'predict' CLI and 'serve' (local POST /v1/systemone server); card: CLI and server docs

Card: zero-shot-classification tag, RAM guide for all variants, limitations, pinned mlx-vlm; clef_mlx: warn on untested mlx-vlm/mlx-lm versions

Add MLX 8-bit conversion of Cloudflare/clef-flash with joint schema head and clef_mlx.py loader

initial commit