idea: memorizing details for top x latency measurements

As an opt-in feature, we could add a ringbuffer per worker for holding more information regarding the requests that saw the top x highest latency replies. This information should be useful for correlating high latency in time, across connections, and pools(workers).

It seems that a way to do this might be fairly simple:

- We take a `std::queue` which will serve as the backing store.
- When a latency measurement is sampled; push it into the queue iff its value is equal
to or higher then the head of the queue
- trim the back of the queue to our desired maximum queue length

Each entry in the ringbuffer stores these fields:

- request time
- reply time
- connection # (if we can easily can get to this with the current state)
- worker nr
- maybe allow storing (a selection of) request/reply information. E.g. it might be useful to
allow storing the value of a selected response field

As it will be known what needs to be stored up front, pre-allocation of memory could be used to avoid allocations in the hot path. Inserts/removals should be fast operations as all operations only involve the front and back of the queue.