fix: stop double-counting partial offload in ranking

[Andyyyy64]

Fixes #105.
Follow-up from #104.

What changed

This fixes a ranking bug where strong models that need partial offload could be buried below weaker full-GPU models.

The old score path counted VRAM fit more than once:

• quality_score already applied the partial-offload quality multiplier
• quality_score also included the speed adjustment
• the final family sort key then added another +15 bonus for full_gpu

That made the displayed score and the actual rank disagree. On an RTX 3060 12GB-style setup, Qwen/Qwen3.6-27B could show a higher score than several models above it, but still land near the bottom because the sort key applied the extra full-GPU bonus.

This PR keeps final sorting closer to the displayed score. Full-GPU candidates are still favored through the runtime-fit and speed terms, but the sort key no longer adds a second full-GPU bonus.

I also made the partial-offload multiplier more granular:

• light spill is penalized less than heavy spill
• heavy dense partial offload is still penalized strongly
• MoE models get a milder penalty when the active parameter working set can plausibly stay on GPU while inactive experts spill

Local QA

uv run pytest
# 331 passed

uv run python -m compileall -q src tests
# OK

Manual checks:

• RTX 3060 12GB: Qwen/Qwen3.6-27B moves from buried near the bottom to the top group with a partial-offload warning.
• A3000 6GB: heavy partial-offload models do not take over the top rank.
• RTX 4060 8GB: Qwen/Qwen3.6-27B stays lower when the offload ratio and speed are not good enough.
• RTX 4090 24GB and M2 16GB rankings still look sane.