Don't buffer postgres sources messages as aggressively

[guswynn]

In #11946, we buffer messages from the pg source in 2 places. We should strive to remove one or both of these:

• We buffer all InProgress in the reclocking process. The ReclockOperator can likely be adjusted to hand out a timestamp but not seal off that timestamp for offsets that have not reached Finalized status cc @cjubb39
• We buffer transactions in the postgres source itself, both in snapshots and replication. These can likely be made to aggressively emit InProgress messages and send Finalized messages at the end

[benesch]

Fixed by #13246.