fix: make the timings recorded in wait_duration and histogram consistent

[aqrln]

Follow-up to #99, there was a comment from me that wasn't addressed: #99 (comment).

We were computing the time spent waiting for a connection twice, once to add it to internals.wait_duration and once to record it in a histogram, so the latter is technically always larger. If the process is preempted by the kernel between these points, the timings can diverge significantly, especially in resource-constrained environments when running many threads/processes.

This PR changes the DurationHistogramGuard::elapsed method into a method that consumes self and records the elapsed time and returns in lockstep, only invoking Instant::elapsed once. The destructor that updates the histogram still exists to make the API misuse-resistant and because that's the contract of a guard, but it's not a code path that happens to be taken anywhere in practice right now.