Lookup joins kept in the island

Applies to: sync and async lookup joins

A lookup join left on the host drags the probe-side Calc and source back to rowwise with it, since the plan splits at the join rather than around it — so both lookup variants are implemented to stay native.

The sync operator keeps probe batches Arrow and calls the connector's real LookupFunction once per row. The point lookup is row-oriented no matter what — the operator around it doesn't need to be.

The async operator fires the connector's asyncLookup for each distinct key in a batch concurrently and joins on the task thread: a batch's lookups overlap, duplicate keys are deduped (safe because the dimension state is fixed within a batch), and all I/O begins and ends inside processElement, so nothing is in flight across a checkpoint. This is the within-batch model Arroyo and RisingWave use, and it avoids Flink's own AsyncWaitOperator mailbox/replay machinery entirely — there is no cross-checkpoint replay queue to manage because no lookup outlives the batch that issued it.