refactor: Respect source transaction boundaries

[chubei]

This PR changes the mechanism of pipeline Commit message generation.

Connector behaviour change

- Connector's "state" used to be defined per table as HashMap<String, TableState>, now it's a single SourceState.

• Connector is expected to send IngestionMessage::TransactionInfo(TransactionInfo::Commit) messages on transaction boundary.

Pipeline behaviour change

Commit used to be generated whenever a source has ingested a certain number of operations, or certain time has passed. Now it's generated when all sources happen to be at a transaction boundary.

Note that if there are multiple sources and data keep flow in, it's possible that pipeline Commit is never generated, because when source A is at transaction boundary, source B isn't necessarily that.

@karolisg I need your review on the postgres connector.
@abcpro1 I need your review on the mysql connector.

[karolisg]

For mysql, I think we need to send a commit on XID_EVENT.

For example

[chubei]

For mysql, I think we need to send a commit on XID_EVENT.

Can you help fix that?

[karolisg]

For mysql, I think we need to send a commit on XID_EVENT.

Can you help fix that?

Sure

[Jesse-Bakker]

In stateless replication, every DAG with be a path, so the case of commits never happening won't occur, right?

[chubei]

In stateless replication, every DAG with be a path, so the case of commits never happening won't occur, right?

Not if we start doing denormalization.