Reuse Parquet OffsetIndex objects and direct-plan decisions #328

Problem

Arrow 17 deserializes OffsetIndex on every lookup, even when filesystem/index bytes are cached. Selective reads repeatedly request the same index during bitmap trimming, prebuffer planning and decoding. Rebuilding direct-read plans also repeats work.

Improvement

Offer reader-local parsed OffsetIndex caching through parquet.read.enable-offset-index-cache, disabled by default to avoid retained-memory growth for wide tables. Reuse the per-leaf direct-plan decision. Preserve page selection, validation and fallbacks.

Evidence

A local 100,000-row, one-column, 1-KiB-value benchmark with 6,250 pages reduced single-row read latency from 2.445 to 0.892 ms with equal bytes read. With 98 pages the improvement was 3.5%; full scans were unchanged. CPU samples confirm repeated OffsetIndex deserialization as the dominant cost. These results support selective many-page workloads, not a general wide-table speedup.

Implementation, reproduction and memory trade-offs: #314.