feat(core): add checkpoint-backed materialization catch-up

[marcus-pousette]

Summary

• add persisted materialization checkpoints on top of the replay-frontier model from PR feat(core): replace dirty materialization with replay frontier #120
• restore the latest checkpoint whose head is strictly before the replay frontier, then replay only the suffix after that checkpoint head
• add key-range op-log scans (scan_after) so Postgres and sqlite can replay from a concrete op-key boundary instead of rescanning from the beginning
• keep external behavior unchanged while reducing the normal catch-up cost for out-of-order persisted remote ops

Why This Is A Separate PR

• PR feat(core): replace dirty materialization with replay frontier #120 fixes the state model by replacing dirty with an explicit replay frontier
• this PR improves the catch-up cost model on top of that new state model
• splitting them lets the architectural cleanup land independently from the checkpoint design

Checkpoint Model

• the op log remains canonical
• materialization metadata still tracks:
  • the materialized head
  • an optional replay frontier
• adapters now also persist historical materialization checkpoints at committed heads:
  • first committed head (seq = 1)
  • then every 64 committed materialization seqs
• catch-up now restores the latest checkpoint whose head is strictly before the replay frontier, then replays only the op-log suffix after that checkpoint head
• if no earlier checkpoint exists, catch-up falls back to replay-from-start

Core Changes

• Storage gained scan_after(...) so adapters can stream ops strictly after a concrete (lamport, replica, counter) boundary
• core catch-up no longer treats replay as “always rebuild from the beginning” in the normal path
• core now asks the backend for:
  • the latest checkpoint before the frontier
  • checkpoint restore into live materialized stores
  • suffix replay after the restored checkpoint head

Postgres

• new checkpoint tables:
  • treecrdt_checkpoints
  • treecrdt_checkpoint_nodes
  • treecrdt_checkpoint_payload
  • treecrdt_checkpoint_oprefs_children
• ensure_materialized_in_tx now restores the latest checkpoint before the frontier and replays only the suffix after that checkpoint head
• successful head updates during append/local/catch-up also persist checkpoints when the seq crosses the checkpoint policy
• Postgres overrides scan_after(...) with a SQL key-range scan over (lamport, replica, counter)

SQLite

• new checkpoint tables:
  • materialization_checkpoints
  • checkpoint_nodes
  • checkpoint_payload
  • checkpoint_oprefs_children
• treecrdt_ensure_materialized() now restores the latest checkpoint before the frontier and replays only the suffix after that checkpoint head
• successful head updates during append/local/catch-up also persist checkpoints when the seq crosses the checkpoint policy
• sqlite overrides scan_after(...) with the equivalent SQL key-range scan

Cases Improved By This PR

• normal out-of-order persisted remote ops no longer imply replay-from-start when a prior checkpoint exists
• catch-up after restart can restore the latest checkpoint before the frontier instead of reconstructing the entire materialized prefix from the beginning of the log
• Postgres and sqlite now share the same checkpointed catch-up behavior

Remaining Fallback

• if a replay frontier predates the earliest retained checkpoint, catch-up still falls back to replay-from-start
• this PR does not implement delta checkpoints or in-place suffix mutation of existing materialized rows; it uses persisted snapshots plus suffix replay

Tests

• cargo test -p treecrdt-core
• cargo test -p treecrdt-postgres --test postgres_test
• cargo test -p treecrdt-sqlite-ext --features ext-sqlite --test extension_roundtrip
• pnpm run clippy

[marcus-pousette]

Superseded by #123, which is the preferred follow-up to #120.