SDK recovery primitive: historical public-note backfill by tag

[haseebrabbani]

Sub-issue of #357. Spike #412 benchmarked the approach and confirmed it is fast without any server-side changes.

Context

Public notes addressed to an account are on-chain, but normal sync starts from the store's global cursor. In a shared dirty store the cursor may already be past blocks containing a recovered account's notes (H < B ≤ T), and a fresh store has no efficient path at all — full syncState() is the "hours" path #357 describes.

Spike findings (#412), testnet at ~1.61M blocks:

• Genesis→tip tag-scoped SyncNotes for one tag: 0.34 s zero-match, 2.0 s at ~3k matching blocks, 44.5 s at 135k matching blocks. Cost scales with matches, not range length.
• For contrast: fresh-store syncState() with that dense tag took 11.8 hours (~950× slower).
• Both the Rust and WASM/TS syncNotes paginate the full requested range internally — no client-side chunking needed.
• Safety cap: miden-client errors (PaginationError) after 1000 pages per tag-chunk rather than truncating.
• The creation-block anchor is therefore optional at current network scale (tracked separately; see the anchor issue) — genesis is an acceptable v1 lower bound and this primitive is unblocked now.

Scope

• TS: backfillPublicNotesByTag(midenClient, { accountId, midenRpcEndpoint, fromBlock = 0, toBlock = tip }) using the raw WASM syncNotes.
• Rust: backfill_public_notes_by_tag(account_id, from_block, to_block) on MultisigClient using the existing sync_notes wrapper.
• Build the account's standard tag (NoteTag::with_account_target), scan the range, keep public notes, fetch bodies/proofs, import individually as NoteWithProof.
• Never mutate the global sync height — normal forward sync runs afterwards.
• Catch PaginationError and fall back to client-side range splitting; report, don't fail recovery.
• PublicBackfillReport { scannedFrom, scannedTo, discovered, outcomes }.

Acceptance criteria

• Both SDKs with semantic parity.
• The dirty-store scenario test: coverage at H, note committed at B, global cursor at T with H < B ≤ T — recovery discovers the note at B without rewinding T.
• Duplicate results and tag collisions are tolerated (tags are best-effort filters).
• Pagination-cap fallback covered by a test.
• Docs: notes sent with unrelated custom tags are outside this guarantee.

Unblocked (spike complete).