Add Phase 2 learned router training and evaluation

[ujjwalkumar26]

Summary

This PR delivers roadmap Part 2 by turning LearnedRouter into a real trainable router and adding the end-to-end pieces needed to use it in practice.

It adds:

• relevance-driven router training
• learned-router evaluation APIs
• persistence via save/load
• public exports for the learned router types
• runnable examples for training and evaluation
• unit coverage for the new behavior
• README updates for training, inference, and evaluation workflows

What Changed

• Reworked LearnedRouter to learn metric weights from labeled relevance data instead of training toward uniform weights.
• Added fit_from_relevance(...) to derive routing targets from per-metric retrieval quality.
• Added evaluate_from_relevance(...) and LearnedRouterEvaluation for measuring learned routing quality.
• Kept the runtime API simple: train a router once, then pass it into refract.search(...).
• Added save() / load() support for trained routers.
• Exported the learned-router classes from the package root and routing package.
• Added examples:
  • examples/train_learned_router.py
  • examples/evaluate_learned_router.py
• Added tests in tests/unit/test_learned_router.py.
• Updated the README to document Phase 2 usage and examples.
• Removed the stale PyTorch requirement from the learned extra, since the new implementation uses existing sklearn-based dependencies.

Notes

• The new evaluation/example flow is designed to help users validate learning on their own data instead of claiming a universal quality lift.
• The sample evaluation on the tiny bundled dataset shows the learned router working end to end; on held-out sample queries it matches heuristic routing, which is expected on such a small corpus.

Validation

Ran locally:

• ruff check src tests
• ruff format --check src tests
• pytest tests/ -x -q --tb=short
• pytest tests/
• python examples/train_learned_router.py
• python examples/evaluate_learned_router.py
• ruff check src tests examples/train_learned_router.py examples/evaluate_learned_router.py
• ruff format --check src tests examples/train_learned_router.py examples/evaluate_learned_router.py

Reviewer Focus

• API shape and ergonomics of LearnedRouter
• training target derivation from per-metric relevance quality
• docs/examples clarity for the new Phase 2 workflow
• package metadata change removing the PyTorch learned extra requirement