Sort README rows by model created_at via cached HF Hub queries

Add an append-only cache at scripts/model_created_at.csv that maps each
HF model referenced from a notebook to its Hub-side created_at timestamp.
The cache is used to order rows within each README section by descending
"most recently created loaded model" so newer model families surface to
the top of their section automatically.

extract_hf_model_refs_from_notebook now returns both the full set of
<org>/<repo> hits and the subset that came from `model_name="..."` or
positional `from_pretrained("...")` calls. The sort key prefers those
"actually loaded" models so notebooks that list alternative variants in a
side block (e.g. gemma4_models = [...]) do not all collapse to the same
key.

The cache is sticky on not_found and only re-queries entries marked
error, so subsequent runs add at most a handful of new repos. HF_TOKEN
is passed when present so gated repos resolve.

Author: danielhanchen