Add idempotent incremental checkpoints and cursor paging

[kraftaa]

Summary

This PR improves incremental sync correctness and rerun safety in rustream by:

• checkpointing progress after each successful batch write
• using composite cursor pagination for duplicate watermark values
• persisting both watermark and cursor in SQLite state
• supporting explicit append-only incremental mode when no tiebreaker is needed
  Why
  Previously incremental sync could skip rows when many records shared the same watermark value (for example identical updated_at). Also, retries after partial failure had weaker resume guarantees.

Changes

1. Incremental checkpointing:

• progress saved batch-by-batch after successful output write

2. Composite cursor pagination:

• query now advances by (watermark, cursor):
• watermark > last_watermark
• or watermark = last_watermark and cursor > last_cursor
• deterministic ordering by watermark and cursor

3. State model:

• watermarks table now includes cursor_value
• backward-compatible migration for existing state DBs
• fail-fast when legacy state has watermark but missing cursor for cursor-based incremental

4. Config updates:

• added tables[].incremental_tiebreaker_column
• added tables[].incremental_column_is_unique for append-only watermark-only mode

5. Safety/validation:

• parameterized watermark query
• validation for missing/invalid incremental/tiebreaker columns
• validation preventing identical watermark and tiebreaker columns

6. Docs/examples:

• updated README.md and config.example.yaml with new incremental config modes

Behavior notes

• For mutable tables, configure both:
  • incremental_column
  • incremental_tiebreaker_column (recommended primary key)
• For append-only unique columns (for example monotonic id), set:
  • incremental_column
  • incremental_column_is_unique: true
• Tables without incremental_column remain full-sync each run.

Testing

cargo test passed locally (44 passed, 0 failed)

Risk / Follow-up

• Existing deployments with old watermark-only state may need state reset for tables moved to cursor-based incremental mode (fail-fast error explains this).