Decouple iceberg commit from risingwave commit.

[liurenjie1024]

See risingwavelabs/rfcs#78

[ZENOTME]

I find that after decoupling, it means that the data is not necessary to flush when it meets a checkpoint. To keep the original semantics of flush, we should need to distinguish it from the regular checkpoints. For regular checkpoint, the sink may not flush the data. But for the checkpoint caused by flush, the sink should guarantee to flush the data.

[liurenjie1024]

cc @wenym1

[ZENOTME]

Following is the modification probably involved:

1. BarrierKind <-- add FlushCheckpoint kind
2. LogWriter::flush_current_epoch <-- add trigger_by_flush param
3. item for LogStoreIml <-- add trigger_by_flush field
   we only need to modify kv_log_store because in_memory log store can't be used in decouple mode.
4. LogStoreReadItem <-- add trigger_by_flush

[liurenjie1024]

Following is the modification probably involved:

1. BarrierKind <-- add FlushCheckpoint kind
2. LogWriter::flush_current_epoch <-- add trigger_by_flush param
3. item for LogStoreIml <-- add trigger_by_flush field
   we only need to modify kv_log_store because in_memory log store can't be used in decouple mode.
4. LogStoreReadItem <-- add trigger_by_flush

LGTM

[ZENOTME]

closed by #15634