perf(index): overlap IVF query partition load and score

[BubbleCal]

Main’s IVF query path already loads partitions one-by-one. The remaining cold-query cost is scheduling: the global-top-k path waits for 64 MiB / 128 partitions before the first score, and prepare concurrency follows CPU count (ncpu - 2), not I/O parallelism.

This PR keeps per-partition loads and adds four independently gated optimizations. Unset LANCE_IVF_QUERY_OPTS enables all of them; none is main’s behavior; a comma-separated list enables one opt at a time.

• first_wave: score after the first prepare wave. Later chunks keep the 64 MiB / 128 bound so a high-nprobes warm query does not pay a dispatch per wave.
• io_prepare: prepare/load concurrency is max(io_parallelism, cpu) (local default 8, not 2 on a 4-core box).
• parallel_files: try_join index.idx and auxiliary.idx in load_partition_entry. Neutral for IVF_RQ (empty index rows); helps IVF_PQ/HNSW.
• overlap_prefilter: start partition I/O without waiting for the prefilter to become ready.

Isolated cold-query results

Open-excluded protocol on local disk: untimed get_ivf_model / centroids, posix_fadvise(DONTNEED) on _indices, then one search. Each row is one opt vs none (main). IOPS, bytes, and recall were identical across every opt.

Mean cold latency (ms), 4-core host, local io_parallelism=8:

Dataset	Index	nprobes	none	first_wave	io_prepare	parallel_files	overlap_prefilter	all
vibe-msmarco-qwen-1024 (8.8M)	IVF_RQ1	20	82	67 (−18%)	61 (−26%)	69	68	58 (−29%)
vibe-msmarco-qwen-1024	IVF_RQ1	80	192	153 (−20%)	134 (−30%)	151	151	141 (−27%)
vibe-msmarco-qwen-1024	IVF_RQ5	20	114	109 (−5%)	108 (−5%)	109	112	104 (−9%)
vibe-msmarco-qwen-1024	IVF_RQ5	80	316	309 (−2%)	268 (−15%)	319	313	271 (−14%)
coyo-ve-qwen3vl-2048 (15.4M)	IVF_RQ1	20	168	160 (−5%)	154 (−8%)	163	162	146 (−13%)
coyo-ve-qwen3vl-2048	IVF_RQ5	20	201	199 (−1%)	206 (+2%)	201	201	194 (−3%)

Takeaways:

• io_prepare is the largest isolated win, especially at nprobes=80 (more in-flight partition reads).
• first_wave helps when the whole probe set is under 64 MiB (RQ1). It does little once a single wave already fills the byte budget (RQ5).
• parallel_files and overlap_prefilter are ~neutral on these IVF_RQ fixtures (empty index.idx rows, no deletion prefilter). They stay in all because they are the right shape for IVF_PQ/HNSW and filtered queries.