feat(snapshot): parallel PK-range chunks with per-chunk resume

[lekhrocks]

What

Reworks the full-table snapshot path into parallel PK-range chunk workers and extracts the JPA persistence layer into a syncflow-persistence module.

Parallel snapshot (F15)

• SnapshotExecutor splits each mapped table into PK ranges and runs up to parallelism worker virtual threads over a shared work queue. Each worker owns one exclusive JDBC connection clone for its lifetime, so no two in-flight tasks ever share a java.sql.Connection (the prior round-robin clone assignment could run two tasks on the same connection).
• Per-chunk resume checkpoints (snapshot_checkpoints.chunk_index, migration V15) — resume continues each chunk from its own keyed cursor. cursorWithinRange guards a legacy whole-table checkpoint from being resumed out-of-range.
• Chunk bounds computed in BigDecimal (AbstractJdbcSnapshotConnector.rangeChunks) to eliminate long overflow on near-full BIGINT domains (max + 1 wrapping, negative rangeSpan collapsing to step = 1). Empirically verified across Long.MIN_VALUE..Long.MAX_VALUE.
• Offset-based sources (Mongo, PK-less JDBC tables) paginate from a per-chunk batch counter instead of the shared global counter, so parallel collections no longer slip or re-read rows.
• Snapshot writes upsert when a destination PK is mapped, so a resume that re-reads already-committed rows is idempotent. Added MySqlWriter.upsertSql (ON DUPLICATE KEY UPDATE) — the base emitted Postgres ON CONFLICT, which a MySQL destination rejected.
• Writer rollback() now discards the residual buffer, so a failed run can't leak buffered rows into the next pipeline's destination.
• Terminal states serialized under one lock: cancel/complete/fail cannot race, and the cancellation flag is not released early (a completing worker can't override a cancelled snapshot).

Multi-tenancy / reliability

• CheckpointStore takes the tenant id explicitly instead of reading the worker-thread ThreadLocal (never set on pool/virtual threads), which keyed every chunk's checkpoints to TenantId.DEFAULT — a second tenant could resume off the first tenant's cursor.
• DistributedLockService guards snapshot start (S2) so two pods can't start the same pipeline concurrently; a stale RUNNING row left by a crashed pod is treated as startable.

Module split

JPA entities, Spring Data repositories, and Flyway migrations (V1–V15) move from syncflow-api into a new syncflow-persistence module (@EntityScan/@EnableJpaRepositories via PersistenceConfig, scanBasePackages = "com.syncflow" picks it up).

Validation

• ./gradlew test — 758 unit tests green.
• ./gradlew test -Dtests.integration=true — 759 tests green incl. SnapshotIntegrationTest (5), ChunkedSnapshotCorrectnessTest (2, real Postgres), DatabaseMigrationValidationTest (11, V1–V15), CdcIntegrationTest (11), KafkaIntegrationTest (4).