Add BM25 ground truth precomputation and validation scripts for the
MS-MARCO v2 (138M) dataset, analogous to the existing v1 pipeline.

Validates 20 curated queries: 10 with high-frequency terms (doc_freq >
8.4M, targeting the block_count overflow bug in #266) and 10 with
low-frequency terms (baseline correctness).

Key design choices for 138M scale:
- Single-pass doc_freq computation (~20 min vs ~13 min/term)
- Materialized doc_term_data and doc_lengths tables to eliminate
  correlated subqueries that block parallel execution
- fieldnorm_quantize marked PARALLEL SAFE for parallel scans
- Total precompute time: ~2 hours (vs ~5+ hours with serial approach)

Validation results:
- Without #266 fix: 4/10 high-freq queries fail (doc mismatches)
- With #266 fix: 20/20 queries pass (worst diff: 0.000002)

Expand validation from 20 curated queries to 400: 10 high-frequency
term queries, 10 low-frequency term queries, and 380 randomly sampled
from the full dev set. Random selection uses hashint4(query_id + 42)
for reproducibility.

All 400 queries validated against the BM25 index with PR #266 fix:
100% docs match, 100% scores match, worst absolute diff 0.000003.

Total precomputation time: ~14.5 hours on 138M rows.

## Summary

- Add BM25 ground truth precomputation and validation scripts for
MS-MARCO v2 (138M docs)
- Validates 20 curated queries: 10 high-frequency terms (targeting #266
block_count overflow) + 10 low-frequency terms
- Includes `ground_truth_pg17.tsv` with precomputed reference scores

## Validation Results

| Condition | Docs Match | Scores Match |
|-----------|-----------|--------------|
| Without #266 fix | 16/20 (80%) | 20/20 (100%) |
| With #266 fix | 20/20 (100%) | 20/20 (100%) |

The 4 failing queries without #266 all contain high-frequency terms
(>8.4M postings) — exactly the terms affected by the uint16 block_count
overflow.

## Design Notes

At 138M rows, naive approaches don't work:
- Materialized `doc_term_data` and `doc_lengths` tables eliminate
correlated subqueries that block parallel execution
- `fieldnorm_quantize()` marked `PARALLEL SAFE` (defaults to UNSAFE,
silently blocks all parallelism)
- Single-pass doc_freq computation instead of per-term COUNT
- Total precompute time: ~2 hours with 13 parallel workers

## Test plan

- [x] Precompute ground truth on PG17 (138M rows)
- [x] Validate against unfixed main — 4/10 high-freq queries fail as
expected
- [x] Validate against #266 fix — all 20 queries pass
- [ ] Review scripts for correctness