Default flow control strategy is not very memory friendly for a cluster with many TiFlash nodes

[windtalker]

Bug Report

In TiDB, it always set the window size to 1G in Grpc client
https://github.com/tikv/client-go/blob/e80e9ca1fe66c8cc251491d880f2d2645293c1a5/internal/client/client.go#L94-L95
It means before the application(TiDB) consume it, the max buffered message inside grpc core can be 1G for each connection.
Consider a cluster with N TiFlash nodes, and a simple query like

select * from t

will dispatched to all the TiFlash node to running parallelly in mpp mode, and at TiDB side, it will establish N connection to these TiFlash node, and receive data from the N connections. If the comsumer speed is slower than the producer(which is almost always true), the data will be buffered in grpc core, and will take up at most N G, if N is larger, it may easily makes TiDB server OOM.