refactor(ranker): split ranking helpers into focused modules

[RaghavRD]

Move ranking orchestration, filters, quality scoring, source trust constants, and candidate variant selection into focused modules.

Keep whichllm.engine.ranker as a compatibility shim so existing public and test imports continue to work.

Refs #41

What

• Move rank_models orchestration into whichllm.engine.ranking
• Move model/profile/evidence filters into ranking_filters
• Move quality scoring and family selection logic into ranking_score
• Move source trust constants into ranking_sources
• Move candidate GGUF/synthetic variant logic into ranking_variants
• Preserve old imports through whichllm.engine.ranker

Why

This continues the #41 core-module refactor by splitting the large ranker module into focused pieces without changing the public import surface.

This is the penultimate major #41 refactor slice. After this, the remaining large core module is src/whichllm/cli.py.

Testing

• Focused ranker/regression tests pass (74 passed)
• CLI monkeypatch compatibility tests pass (6 passed)
• Ruff check passes
• Ruff format check passes
• Full suite except environment-sensitive run smoke test passes (465 passed, 1 deselected)

Notes

Full pytest in my sandbox hit test_run_exits_gracefully because model fetching returned a DNS/network error message from the environment. The ranker-specific and compatibility coverage passes.