chore: dont use loop on yamux

[driftluo]

Rust's asynchronous system is a collaborative model, which means that all Futures should not occupy the executor for a long time.

There is a risk of preemption when yamux uses a loop. There will be no obvious problems under a single connection. However, when a large number of connections exist, if yamux is occupied for a long time, it will affect the scheduling of the entire task system, especially when the CPU capacity is insufficient. It will become more and more obvious on the machine, the consequence is that the communication speed will be greatly jittered.

What are the problems with occupying the executor for a long time?

First, it will cause the work-steal trigger to become more frequent
Second, it will affect the drivers triggering

Generally, Future tasks use executor in the principle of proximity, and bind on an executor, triggering work-steal means that there is additional synchronization work

Here we specify the maximum upper limit of one poll. If yamux still needs to be executed, first yield out, perform other tasks, or execute itself again

ref: tokio-rs/tokio#3493