Graph-IVF: add SPFresh-style LIRE maintenance

[xwj-ox]

• Does this PR have a descriptive title that could go in our release notes?
• Does this PR add any new dependencies?
• Does this PR modify any existing APIs?
• Is the change to the API backwards compatible?
• Should this result in any changes to our documentation, either updating existing docs or adding new ones?

Reference Issues/PRs

SPFresh: https://www.microsoft.com/en-us/research/publication/spfresh-incremental-in-place-update-for-billion-scale-vector-search/

What does this implement/fix? Briefly explain your changes.

This replaces Graph-IVF's independent split and dissolve maintenance with a synchronous, in-memory adaptation of SPFresh's LIRE protocol. It adds capacity-balanced binary splits, the two necessary-condition filters for possible NPA violations, deduplicated global routing of filtered candidates, post-retirement merge rerouting, and split/merge cascade convergence.

The merge planner first searches the existing mutable centroid Vamana graph, exactly reranks the bounded candidate set, and applies projected-capacity filtering. It progressively widens the candidate budget and uses a packed, one-pass exact fallback without restoring the previous all-centroid allocation and full sort. All victims in the same batch are excluded, and planned target capacity is reserved across the batch.

The change preserves final global NPA routing and overflow cascades. It does not implement SPFresh's asynchronous Local Rebuilder, version/replica GC, SSD Block Controller, or snapshot/WAL system layer.

Validation completed locally:

• cargo test -p diskann-graphivf: 89 unit tests and 11 integration tests passed.
• cargo clippy -p diskann-graphivf --all-targets -- -D warnings passed.
• cargo fmt --all --check and git diff --check passed.
• A preliminary MSTuring30M replay preserving all 29,998,834 inserts and 27,630,980 deletes reduced LIRE merge-target search from 8,733.0 s to 46.6 s and complete merge maintenance from 9,115.4 s to 551.5 s, while merge/split counts, final cluster count, posting equilibrium, and matched-scan recall remained comparable. That run sampled read-only search checkpoints and is not the final stability result.

Any other comments?

This is intentionally a draft. A fresh A1-B1-B2-A2 comparison against u/adkrishnan/graph-ivf is running with the original 1,280-stage MSTuring30M runbook, all 640 search checkpoints, 1,000 queries per checkpoint, per-run corpus warmup, isolated indexes, and interleaved final disk sweeps. The PR should remain draft until the ABBA variance analysis is complete and the clusters_updated instrumentation from 90cb5caf is reconciled across both arms. Local experiment logs and HTML reports are intentionally excluded from this PR.