Optimize the snapshot read of CDC backfill

[StrikeW]

Existing implementation of cdc backfill will start a new snapshot upon receiving a barrier message. And to ensure make progress of the backfill, we guarantee to consume and emit a snapshot chunk for each epoch, which could cause barrier piles up in some scenario for example the wide table use case.

As comparison, here is the metric that we alter the barrier interval to 5 mins instead of apply the patch #15684, the barrier doesn't pile up.