Add reproducible conversational retrieval baselines and paired CIs

[bigboateng]

Summary

Adds paired conversational-retrieval evaluation, reproducible profiles, and an opt-in Gemini rewrite baseline for TREC CAsT.

Evaluation and baselines

• Preserves conversation IDs and adds paired nDCG@3 delta intervals against pointwise retrieval
• Pins the TopiOCQA n=25 MiniLM/kNN profile and committed result
• Adds gemini_rewrite using GEMINI_API_KEY, gemini-3.5-flash, and minimal thinking
• Records provider/model/prompt provenance and compares with CAsT's manual rewrite oracle
• Adds CAsT conversation limiting and scripts/reproduce_cast2019_gemini.sh

Real-data scalability

• Installs TREC CAR support and prevents failed 38M-document builds from restarting per passage
• Uses blockwise NumPy top-k graph construction
• Bypasses graph searches for β=0
• Uses SciPy compiled sparse shortest paths with compact bounded distance arrays and a standard-library fallback
• Vectorizes full-corpus retrieval scoring and bounds the query score cache

CI compatibility

• Replaces unreachable action commits with actions/checkout@v7 and actions/setup-python@v6
• Restores Python 3.11 NumPy compatibility and correct base-package test scope

TopiOCQA result

• β=0.5 follow-up delta: +0.045, paired 95% CI [+0.018, +0.077]
• β=1.0 follow-up delta: +0.084, paired 95% CI [+0.046, +0.128]
• β=1.0 standalone delta: +0.031, paired 95% CI [+0.009, +0.055]
• MMR follow-up delta: +0.001, paired 95% CI [+0.000, +0.003]

Test plan

• 83 tests passed, 1 skipped before compact sparse caching
• Ruff, formatting, strict mypy, package build
• TopiOCQA profile reproduced twice with byte-identical output
• CAsT combined MS MARCO/TREC CAR docstore built successfully
• Verify compact sparse caching and complete the CAsT Gemini profile
• Final GitHub Actions rerun