Tracking Issue: Skipping Indexes on `ZonedLayout`

[connortsui20]

This is a tracking issue for adding skipping indexes on top of the existing zoned layout, using the per-zone summary machinery already in vortex-layout (layouts/zoned) and vortex-array (aggregate_fn, stats/rewrite).

Motivation

Parent epic: #8900

Design

ZonedLayout already pairs a data child with a per-zone stats table and prunes by running Expression::falsify(predicate) over that table (ZoneMap::prune), then expanding the per-zone skip mask in ZonedReader::pruning_evaluation. This should be sufficient for a skipping index.

Adding an index on this path is two registrations, plus a third only when the summary needs custom test logic:

1. A serializable aggregate for the per-zone summary (AggregateFnVTable).
2. A StatsRewriteRule that lowers the query predicate into a proof over the summary.
3. If the test is not expressible with builtin comparisons (e.g. a bloom membership test), a probe scalar function (ScalarFnVTable) that the rule applies to the stored summary. Min/max-style indexes don't need this.

Scans consume these indexes automatically: the pushed-down filter drives ZonedReader::pruning_evaluation in the existing prune/filter/project loop, so no query plan changes are needed.

Note that on read, an unknown aggregate ID currently hard-errors (zoned/schema.rs::to_aggregate_fn, no allows_unknown fallback, unlike aggregate_fn/proto.rs::from_proto), so a custom index is currently a hard read-dependency rather than an ignorable hint.

A later phase should bundle these registrations behind a single interface so that adding a skipping index is one implementation instead of several coordinated registrations that are easy to half-wire (store the aggregate but forget the rewrite rule, and it silently prunes nothing). The exact shape (one trait vs. composable seams) is the epic's open interface question (#8900); this issue tracks implementing whatever is chosen on top of the zoned layout.

Steps

• Document the extension path (aggregate + probe + rewrite rule).
• Resolve ignorability: wire allows_unknown/foreign into the zoned read path, or require ignorable third-party indexes to be their own layout: Bypass unknown aggregates in ZonedLayout #8904
• Land a proof-of-concept index (bloom filter) with a roundtrip test.
• Design and build an ergonomic interface that bundles the index registrations into a single definition.
• Documentation

Unresolved questions

• How small is wiring allows_unknown/foreign into zoned/schema.rs::to_aggregate_fn, and what state dtype would a foreign aggregate present?
• Is there any summary whose prune cannot be expressed as a falsify rewrite, and thus needs a different hook?
• Does the write-time "index on column X" declaration surface belong here or in its own issue?
• What does the bundled interface look like concretely, and does the raw path stay as an escape hatch for indexes that don't fit it?

Implementation history

• Bypass unknown aggregates in ZonedLayout #8904