feat: add opt-in V3 delta search refresh

[skyoo2003]

Pull Request

Description

Updating one keyword in a large V3 dictionary currently rebuilds the entire search automaton before WaitForVersion completes. This adds an opt-in MemoryEfficient path that downloads changed buckets, publishes a base automaton with a small additions automaton and deletion set, and compacts it after one idle second. Changes above 128 keywords and other presets retain the full rebuild path. The Redis schema and V1/V2 behavior are unchanged.

The option remains disabled by default because the release gate has not been measured: median commit-to-visibility latency for one add and one removal must each improve by at least 50% across Redis and Valkey and three million-keyword distributions, while search p95 and peak RSS each worsen by no more than 20%. The opt-in scale test now records commit-to-visibility timing and repeated single-keyword updates.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Refactoring (no functional changes)
• Test update

Checklist

• Tests pass (make test)
• Vet/make vet
• Linting passes (make lint)
• Build succeeds (make build)
• Documentation updated if needed
• Changelog fragment added (changie new) if this change belongs in the changelog — one sentence (two at the very most), on a single line, 400 characters max; skip for internal-only changes (CI, tests, refactors); see RELEASE.md
• Commit messages follow guidelines

Additional Notes

make all passed, including API audit, documentation verification, and license check. go test -race ./pkg/acor -run TestVersionedOverlayLifecycle -count=1 passed. Million-keyword Redis/Valkey performance validation remains outstanding; the new path is opt-in until that gate passes.

---

By submitting this PR, I agree that my contributions will be licensed under the Apache License 2.0.