feat(index): read a column's min/max from ZoneMap without a scan

[Ali2Arslan]

Summary

A ZoneMap scalar index already stores per-zone min/max summaries. This PR folds those summaries into a single global (min, max) for a column without scanning any data, and exposes it on the dataset:

• ZoneMapIndex::value_range() / ZoneMapIndex::value_range_over(segments) (lance-index) — fold one or more ZoneMap segments.
• DatasetIndexExt::zonemap_value_range(column) (lance) — the dataset-level accessor.

This is useful for cheap min/max stats and as a planning input (e.g. range pruning) without paying for a scan.

Soundness

The folded range is designed to be safe to prune with — it never drops a matching row:

• NaN bail. ScalarValue's total order ranks NaN above every finite value, so a NaN-bearing zone records max = NaN, hiding its true finite max. Folding only the finite maxes would yield a subset that could prune live rows, so any NaN zone makes the result None.
• Joint fragment coverage. The dataset accessor returns None unless the column's ZoneMap segments jointly cover every live fragment. Fragments appended after the index was built (or a segment set that doesn't span the table) leave a live fragment uncovered → None. Extra dead fragments in the union are fine.
• The disjoint segments of a multi-segment index are folded together.
• The result is a superset of live values, conservative under deletion vectors (a deleted extreme still bounds its zone): not guaranteed tight.

Reuses the existing scalar_is_nan helper rather than duplicating NaN logic.

Test plan

lance-index (value_range / value_range_over):

• spans multiple fragments/zones
• all-null → None
• NaN max → None
• folds across segments; NaN in any segment → None; skips all-null segment

lance (zonemap_value_range):

• basic range; column without a ZoneMap → None

• None when an appended fragment isn't covered by the index

• folds multiple committed segments of one logical index

• cargo fmt --all

• cargo clippy -p lance-index -p lance --tests clean

Follow-up

Python bindings (Dataset-level accessor) can follow in a separate PR to keep this one focused on the Rust core + dataset API.

Made with Cursor