perf(index): window IVF partition loads during query

[BubbleCal]

What is the performance issue?

PR #9049 made IVF prewarm 30x faster by reading adjacent partitions in 64 MiB windows instead of one read_stream per partition. Query-time loading was left on the old per-partition path. A cold search still issues a decoder setup, page-metadata init, and object-store GET round for every probed partition.

Query nprobes are usually scattered, so a naive copy of prewarm's contiguous windows would almost never fire. The I/O scheduler only coalesces ranges that arrive in the same submit_request, so concurrent single-partition reads do not merge even when pages are nearby.

How does this PR improve performance?

Reuse the prewarm reader for query-time loads:

• Uncached partitions in a prepare-window chunk (prepare_parallelism, typically the CPU pool size) are sorted and grouped by the same encoded-byte / partition caps as prewarm (LANCE_IVF_PREWARM_WINDOW_SIZE_BYTES, default 64 MiB).
• Adjacent IDs with a continuous IVF row layout use one Range. Scattered nprobes use one Ranges request so page-metadata init and nearby GETs can coalesce.
• Single-flight cache inserts stay per-partition (leader + followers), matching prewarm.
• Search still streams those chunks into the existing global-top-k / streaming scoring batches, so resident prepared-partition memory stays bounded by the prepare window plus scoring chunks — not by nprobes.

load_partition (single ID) is unchanged for search_in_partition and other one-partition callers.

HF retrieval benchmarks (open-index excluded)

4 vCPU / 15 GiB, local disk, pylance 13.0.0-beta.4. Cosine IVF_RQ, no refine, k=100.

Protocol (cold sample): open the dataset and call get_ivf_model / centroids outside the timer (this is 8 IOPS / ~4 MiB). posix_fadvise(DONTNEED) on _indices only. Reset I/O counters. Time one search on that already-open handle. Fresh Session per cold sample so the partition cache is empty. Warm = one session, 5 untimed queries, then measure.

windowed is the default 64 MiB grouper. singleton sets LANCE_IVF_PREWARM_WINDOW_SIZE_BYTES=1 (same new reader, one partition per group).

Build (centroids reused for RQ5):

Dataset	Rows	Dim	nlist	RQ1	RQ5
vibe-msmarco-qwen-1024	8,840,823	1024	1024	393s / 1.3 GiB	416s / 5.6 GiB
coyo-ve-qwen3vl-2048	15,380,795	2048	4096	1030s / 4.0 GiB	1958s / 19 GiB

MS MARCO (8.8M × 1024-d, 20 queries)

Index	nprobes	Mode	Cold lat	Cold IOPS	Cold MiB	Warm lat	Recall@100
IVF_RQ1	20	windowed	200 ms	214	26	34 ms	0.794
IVF_RQ1	20	singleton	29 ms	217	26	19 ms	0.794
IVF_RQ1	80	windowed	205 ms	512	102	49 ms	0.824
IVF_RQ1	80	singleton	88 ms	518	102	32 ms	0.824
IVF_RQ5	20	windowed	442 ms	273	116	104 ms	0.905
IVF_RQ5	20	singleton	75 ms	278	116	39 ms	0.905
IVF_RQ5	80	windowed	601 ms	750	461	138 ms	0.966
IVF_RQ5	80	singleton	260 ms	758	461	77 ms	0.966

Coyo-VE (15.4M × 2048-d, 15 queries)

Index	nprobes	Mode	Cold lat	Cold IOPS	Cold MiB	Warm lat	Recall@100
IVF_RQ1	20	windowed	186 ms	228	23	19 ms	0.852
IVF_RQ1	20	singleton	35 ms	232	23	18 ms	0.852
IVF_RQ1	80	windowed	282 ms	527	87	43 ms	0.887
IVF_RQ1	80	singleton	96 ms	532	87	36 ms	0.887
IVF_RQ5	20	windowed	523 ms	286	107	58 ms	0.920
IVF_RQ5	20	singleton	112 ms	292	107	52 ms	0.920
IVF_RQ5	80	windowed	1153 ms	764	412	156 ms	0.971
IVF_RQ5	80	singleton	329 ms	772	412	129 ms	0.971

What this shows

• Opening the index is cheap (8 IOPS / 4 MiB) and is no longer in the timed region. The ~210–770 query IOPS are partition loads: about 10 IOPS per probed partition across index.idx + auxiliary.idx.
• Bytes and recall are identical windowed vs singleton.
• IOPS drop only ~1–2% with grouping. Scattered nprobes do not become prewarm-style sequential windows.
• On this local disk, the 64 MiB grouper is slower for cold latency (often 3–7× vs singleton): same bytes, extra decode / larger materialize. Singleton is the fair picture of today's per-partition load cost (RQ1@20 ≈ 29–35 ms, RQ5@80 ≈ 260–330 ms).
• The perf(index): read IVF prewarm in parallel byte windows #9049 win came from contiguous partition ranges on S3. That shape is not typical nprobe sets.

Tests

• test_plan_query_load_groups_* and test_selected_partition_row_span_* cover sparse grouping, byte/partition caps, and contiguous vs gapped row spans.
• test_query_partition_load_uses_windowed_io builds a 16-partition IVF_PQ index, compares grouped-load IOPS to a single-partition load (must be a small multiple, not N×), checks concurrent load_partitions does not deadlock, and asserts a follow-up query issues zero I/O.
• Existing prewarm and test_global_topk_search_bounds_in_flight_prepared_partitions still pass.

Lint: cargo fmt --all and cargo clippy -p lance --tests --benches -- -D warnings passed. Full-workspace clippy was not rerun; only rust/lance/src/index/vector/ivf/v2.rs changed.

cargo test -p lance --lib -- index::vector::ivf::v2::: 116 passed.