Sort README rows by HF Hub popularity (downloads + likes*1000)

Switch from "most recently created loaded model" to a popularity score
because the previous timestamp signal got distorted whenever unsloth
re-uploaded a quantized variant (e.g. Llama 3.1's Feb 2025 re-upload
pushed it above Llama 3.3's December 2024 release in the section).

Cache schema: scripts/model_created_at.csv now carries downloads, likes,
fetched_at, status alongside created_at. Older CSVs without the new
columns load with downloads=0 likes=0, which the next refresh fills in.

Refresh policy:
  * Missing repo: always fetch.
  * status=error: always retry.
  * status=ok and fetched_at older than 7 days: refresh popularity counts.
  * status=ok and fresh: skip.
  * status=not_found: sticky.

The sort key per notebook is the max popularity score across the models
the notebook actually loads (model_name="..." or positional
from_pretrained), with fallbacks to the broader ref set, then
version-from-name, then row text. Notebooks with no resolvable model
get score 0 and sort to the bottom of their section.

Author: danielhanchen