[DO NOT MERGE] docs: add DiskANN disk-backed index engine SEP

[swasik]

This PR is used just to make the initial review of SEP. To be moved to Confluence later.

What

Adds a ScyllaDB Enhancement Proposal (SEP) for a DiskANN (SSD-backed) index
engine in Vector Store, offered as an additional, selectable engine alongside
the existing in-memory USearch engine.

Docs-only change:

• docs/dev/diskann/sep-diskann.md — the proposal.

Why

Vector Store builds every ANN index fully in RAM (USearch/HNSW), so RAM is the
scaling ceiling and the primary cost driver. DiskANN moves the bulk of the index
to local NVMe SSD (keeping a PQ-compressed representation + working set in RAM),
targeting:

• Large workloads (100M–1B+): materially lower $/indexed-vector and
  $/QPS at acceptable recall, trading a modest latency increase for a large
  RAM-cost reduction.
• Cost-sensitive smaller workloads: a lower cost floor — run on cheaper
  storage-optimized nodes when higher latency is acceptable.

USearch remains the default for latency-critical, RAM-resident workloads.

Design highlights

• New DiskannIndexFactory behind the existing IndexFactory seam (the same
  pluggable-engine plumbing USearch uses); ingestion (CDC), routing, and the
  HTTP/CQL surface are unchanged.
• Integrates DiskANN3 (Microsoft Research, Rust, MIT, v0.54.0) in
  disk-provider mode; node-local NVMe to avoid a network hop per beam-search
  step.
• Real-time updates via IP-DiskANN (in-place, natively supported by
  DiskANN3) — no periodic merges or full rebuilds.
• Engine selected per node (memory-optimized vs storage-optimized), surfaced
  in Cloud as a single latency-optimized vs cost-optimized deployment switch.
• Durability boundary is the SSD volume; ScyllaDB remains the authoritative
  source, so index loss = rebuild-from-CDC.

Scope

• In scope of the SEP (later milestone): filtered ANN parity with USearch.
• Out of scope: replacing USearch, non-disk DiskANN3 providers, S3/remote
  object storage on the hot path, DistributedANN, GPU, Optane/PMem.

Notes for reviewers

• This is a Draft SEP; several [Add link] placeholders and an Open
  Questions list (PQ defaults, cache-vs-FTS RAM split, p99 target validation,
  crash-consistency strategy) are intentionally left for delivery-time
  resolution.
• Benchmark plan compares USearch (f32, i8, b1) vs DiskANN at 100M and 1B,
  with cost ($/indexed-vector, $/QPS) as a first-class result.

[Copilot]

Pull request overview

Adds a draft ScyllaDB Enhancement Proposal (SEP) describing a DiskANN (SSD-backed) index engine option for Vector Store, positioned as a capacity/cost-optimized alternative to the existing in-memory USearch engine.

Changes:

• Introduces a new draft SEP detailing product motivation, goals, and competitive positioning for a DiskANN-based engine.
• Documents a proposed high-level integration plan (factory selection, actor wiring, config/metrics/memory accounting).
• Outlines QA/benchmark strategy, delivery milestones, and operational considerations for an SSD-backed tier.