Add Multihop Wikipedia QA project

[abderrahim-lectures]

Closes #285

Adds the Build a Multi-Hop Question-Answering Tool Over a Small Wikipedia Sample project.

What's included

• Lesson (docs/projects/multihop-wikipedia-qa/index.md): two-round (iterative retrieval) RAG — single-hop baseline vs. multi-hop pipeline over a bundled Wikipedia-style sample, with an evidence-audited side-by-side comparison. Steps, checklists, Socratic questions, pitfalls, related projects.
• Example (examples/multihop-wikipedia-qa/): six crafted articles, six test questions (three genuinely multi-hop), main.py CLI (--rebuild/--query/--provider/--top-k), Colab/Kaggle/Binder notebook, README, pyproject + uv.lock, .env.example, .gitignore, .python-version.
• Registrations: docs/projects/index.mdx, src/data/projects.ts, src/pages/index.tsx.

Verification

• npm run typecheck ✅
• npm run build ✅ (only pre-existing docs-qa-bot anchor warnings, untouched)
• Offline smoke test: index build (20 chunks, 384-dim), retrieval spot-checks confirm the multi-hop premise (missing facts NOT in top-3 for single-hop; follow-up queries surface them), mock-LLM control flow test of both SUFFICIENT and INSUFFICIENT branches ✅
• Limitation: no real LLM API key available in this environment — the final single-vs-multi end-to-end run on a live provider was not executed. The notebook's last cell was run with an invalid key and failed only at the API call (expected).