Optimization of arctic dim table join in the Flink engine

[YesOrNo828]

According to the Flink Dim performance test guide,
The left table QPS: 1w/s
the right table size: 10G, 1580w
the parallelism is: 8
The first-time join match cost 3mins, which is still too high.
We should consider reducing the cost.

[YesOrNo828]

There are some thinkings about it.
Maybe we can add a flink planner rule and an enforced-stop-reading operator chained after the left table scan source operator.
Simply Physical Graph:

JOIN
  |---------------------|
  |                     |
Enforce-reader          |
  |---------------------|
  |                     |
Left Source        Right Source

Enforce-reader operator has two input streams and one output stream.

Step1: Right Source send a msg start-load-into-rocksdb-event to Enforce Read, and Enforce reader stop reading the data from Left Souce;
Step2: Right Source emits the dim table data to the JOIN operator to initiate data into the state;
Step3: When Right Source completes the historical data reading, send a msg end-load-into-rocksdb-event to Enforce read, and the Enforce read operator begins to read data from the Left Source;
Step4: Join operator begins to match left and right streaming data and emit the result to the next downstream.