estimate the memory consuming spike for the coprocessor

[Yisaer]

Feature Request

Is your feature request related to a problem? Please describe:

In this request #21168, we use ratelimit to limit the max count of cached response size in the tidb-server for one query. The coprocessor is response for the fetching data from storage and works as a producer/consumer model. There exists multiple workers(as producer) try to fetch data from storage and only one Iterator(as consumer) try to receive data and deliver them to the upon operator.

The initial count of workers is depended on the tidb_distsql_scan_concurrency(default 15) which may cause the memory consuming spike at the beginning for the coprocessor.

Describe the feature you'd like:

Find a way to estimate the memory consuming spike for the coprocessor. For example, we could modify the count of the workers with a proper count at the beginning.

Note that there could exist multiple coprocessor operator for one query, we should also consider and test this case.

Describe alternatives you've considered:

Teachability, Documentation, Adoption, Migration Strategy: