Resolve base_model upstream when scoring popularity

Sorting by the unsloth quantized variant's downloads alone gave a
distorted picture: e.g. unsloth/Qwen2.5-VL-7B-Instruct-bnb-4bit only has
~13k downloads while the canonical Qwen/Qwen2.5-VL-7B-Instruct has ~4.4M.
The Vision section ended up with LFM2.5-VL-1.6B at the top because LFM
publishes only one repo (no quantized siblings to dilute its score)
while every Qwen/Llama vision notebook scored against the unsloth quant.

Fix: when fetching model_info, also extract card_data.base_model and
store it on the cache row. After the notebook-referenced fetch pass,
run a second pass that pulls in any base_model upstream not already in
the cache. _popularity_score then follows base_model and returns the
MAX of (self, upstream), so the unsloth notebook inherits the popularity
of its canonical upstream model.

Cache schema gains a base_model column. Older CSVs load with base_model="".

Vision section now sorts as Qwen2.5 VL > Qwen3 VL > Qwen3.5 > Gemma3
GRPO > ... > LFM2.5 VL, matching the actual model family popularity on
the Hub.

Author: danielhanchen