perf(index): bulk conjunction path for FTS AND and phrase queries

[BubbleCal]

Stacked on #7604 (base of this PR); part of the Lucene-parity series after #7600-#7605.

What

Top-k AND and phrase queries previously leapfrogged doc-at-a-time through boxed PostingIterator::next calls — 61% of the AND profile went to per-doc advance machinery (~25-40ns per advance). Phrase checks additionally decoded a whole 256-doc position block per candidate (39% of the phrase profile) and allocated cursor vectors per candidate.

This adds a bulk conjunction path (and_bulk_search, default on, LANCE_FTS_BULK_AND=0 opts back into the classic loop):

• Same block-max window pruning as the classic loop, but candidates come from a k-pointer merge over the decompressed block slices — the window ends at the nearest next-block boundary across clauses, so each clause contributes exactly one block per window and the merge runs on plain u32 slices.
• Two-pass batched scoring: the merge only records (doc, per-clause offsets); doc lengths are gathered back-to-back (cache misses overlap) and the prune/verify/score/insert pass replays the classic loop's exact semantics.
• seek_packed_doc_positions: PackedDelta position groups are self-describing ([num_bits u8][16*num_bits bytes]), so group offsets are recovered by hopping headers — a phrase candidate decodes only the 1-2 groups overlapping its own doc instead of the whole block, with a lazily-built group index and a decoded-tail cache. No format change.
• check_exact_positions_bulk: allocation-free slop=0 alignment check over the decoded scratch slices.
• The doc-length gather goes through DocSet::scoring_num_tokens, so V3 partitions score with the quantized lengths feat(fts)!: add configurable posting block size #7466 defines. (The impact-score-cache dead-code cleanup originally carried here moved to perf(fts): bulk MAXSCORE search path for top-k disjunctions #7603, where the code actually becomes dead.)

Results (mmlb-200m warm, 8 concurrent, vs Lucene 10.4)

query	before	after
AND k10 @200M	0.114s / 69 qps	0.060s / 132 qps
AND k100 @200M	0.240s / 33 qps	0.118s / 67 qps
phrase 3w k10 @50m	0.335s / 23.8 qps	0.221s / 36.1 qps
phrase 2w k10 @50m	0.098s / 81.6 qps	0.048s / 165 qps

Verification

• Results are identical to the classic loop: bulk-vs-classic A/B on 200 query×k pairs, score_diff=0 (AND @200M and phrase @50m).
• New rstest parity suite runs both paths on multi-block corpora (2/3/4 clauses × AND/phrase × k) and asserts identical candidates; encoding roundtrip test covers per-doc seek vs whole-block decode for every doc across tail/group-straddling shapes.
• cargo test -p lance-index, clippy -D warnings, fmt --check clean.

🤖 Generated with Claude Code