feature: incrementally consolidate and backfill column statistics

Parent: #8461

Goal

Build and maintain consolidated column statistics from write-time ZoneMap seeds, scanning only data that lacks usable statistics. Updating statistics should not require data compaction.

Existing foundation

#7427 added seed harvesting during index updates. #6593 and #7128 provide ZoneMap segments and merging. The current seed-harvesting path is all-or-nothing: missing seed data can cause the update to scan all unindexed data. This work should support mixed seed and scan inputs incrementally.

Scope

• Backfill existing fragments when statistics are first enabled.
• Harvest compatible seeds and scan only the uncovered fragments/columns that require backfill.
• Create, refresh, and merge central index segments through the existing index APIs.
• Define automatic maintenance triggers, thresholds, and bounded resource use, with an explicit invocation path for callers.
• Commit statistics together with accurate coverage for the intended dataset snapshot, respecting concurrent writes, schema changes, and index drops.

Write-time seed collection and column policy belong to #4522. Read-side coverage resolution and the public statistics API are separate children of #8461. Ordinary statistics retrieval must not silently start this maintenance work.

Acceptance criteria

• Existing datasets can be backfilled without requiring compaction.
• An update combines compatible seeds with scans of only missing data.
• Fully seeded input updates statistics without scanning source column values.
• Incremental refresh and segment merging avoid double counting and preserve per-column coverage.
• Document trigger policy, resource limits, and retry behavior.
• Concurrent append, delete, rewrite, schema changes, and index drop cannot publish incorrect coverage or restore a dropped index.
• Tests cover multiple fragments and segments, mixed seed availability, incompatible seeds, and interrupted/retried work.
• Record scan I/O, seed-read I/O, refresh throughput, and peak memory in the column-statistics benchmark.

Start with parent issue #8461 and the existing index APIs described in #7427, #6593, and #7128 to understand seed harvesting, ZoneMap segments, and merging. Use the column-statistics benchmark as the entry point for measuring scan I/O, seed-read I/O, throughput, and memory. Done means incremental backfill and refresh handle mixed or incompatible seeds, concurrent changes, retries, and index drops while meeting the listed acceptance tests.