[MPP worker] optimize exchange operator to fully use network throughput

Why

The network utilization decides the performance of MPP, we should carefully design and implement the exchange operator.

What

Optimize the exchange operator to fully use the network throughtput. Like KRPC 实现 Impala 的飞跃, we should let the RPC framework can scale out to many nodes.

How

efficiently multiplex resources: build almost one TCP connection among any two nodes using one thread;
support high concurrency when the computation part reads/writes from/into the data buffer in the exchange operator;
support data compression.

Check

building a test to collect the cost time of shuffle, as well as recoding the network

Start by locating the MPP worker exchange operator and the shuffle path described in the issue. Define and run a benchmark that measures shuffle time and records network utilization; completion requires evaluating multiplexed node connections, concurrent buffer reads and writes, and data compression against network throughput.