perf: insert tail latency issue

[MrCroxx]

Sometimes insert tail latency is abnormally high. Needs investigation.

[MrCroxx]

A case that rotates a few times leads to insert slow. (And rotate is slower than expected, too.)

[MrCroxx]

Sometimes leader acquire clean buffer can be slow. That may hold back followers.

[MrCroxx]

The slow "acquire clean buffer" is blocked at "wait notification", which is notified when a new clean buffer is released.

[MrCroxx]

For rotate tail latency, which is caused by clean buffer acquiring, the root cause is wirte buffer is exhausted and waiting for flushing.

But for allocate tail latency, which is caused by multiple rotating, needs to be optimizated by introducing ordered queue to prevent from starving.

[MrCroxx]

close via #184 then #230