feat: add direct rewind materialization fast path

[marcus-pousette]

Summary

• build on #120's frontier model with a direct rewind/replay fast path for common out-of-order insert / move / payload appends
• keep delete / tombstone on the conservative replay-from-frontier path for now
• add shared Postgres/sqlite conformance coverage for both the fast path and the conservative fallback path

Why

After #120, materialization state is explicit (head + replay_from) and out-of-order remote ops no longer use a coarse dirty flag.

This PR improves the normal catch-up path for common out-of-order sync cases.

Instead of deferring those cases to a later read and replaying broadly from the op log, append-time catch-up now:

1. records the earliest invalidation frontier
2. scans the invalidated suffix in canonical op-key order
3. rewinds the already-materialized suffix directly on backend stores
4. truncates suffix oprefs_children
5. replays the invalidated suffix forward directly on backend stores
6. clears the frontier when complete

What Changes

For common out-of-order insert / move / payload appends, the normal path no longer relies on:

• read-time catch-up as the primary repair path
• full materialized-table reset
• full op-log replay
• full in-memory clone of the materialized doc

A narrow payload shortcut also remains in place for dominated out-of-order payload ops that do not change materialized payload state.

Fallback Behavior

This PR is intentionally conservative for the annoying cases:

• if the invalidated suffix contains delete or tombstone, append falls back to the existing replay-from-frontier catch-up path
• deferred ensure_materialized recovery still uses the conservative replay path

So the fast path is targeted at the common structural/payload out-of-order cases, while delete/revival semantics stay on the safer route.

Backend Changes

Postgres

• add ordered frontier-range scans and per-node predecessor lookups in PgOpStorage
• use the direct rewind/replay fast path during append-time immediate catch-up

sqlite

• add the same frontier-range scans and per-node predecessor lookups in SqliteOpStorage
• use the direct rewind/replay fast path during append-time immediate catch-up

Tests

Verified locally with:

• cargo test -p treecrdt-core --test materialization_helpers
• cargo test -p treecrdt-postgres --test postgres_test
• cargo test -p treecrdt-sqlite-ext --features ext-sqlite --test extension_roundtrip
• pnpm run clippy
• pnpm -C examples/playground typecheck
• pnpm -C examples/playground build

The shared backend conformance suite covers:

• common out-of-order append immediate catch-up
• dominated out-of-order payload shortcut behavior
• replay-from-start frontier catch-up
• deferred ensure_materialized recovery
• explicit delete/tombstone fallback behavior
• rollback behavior when immediate catch-up fails