Fix a data race in a metrics registry without serialising every read

A latency aggregator shared by every request goroutine and one exporter goroutine fails under the race detector. A mutex was added around the map at some point and the work stopped there: the per endpoint aggregate fields, the histogram behind them and the global Observe counter are all mutated with no lock held, and a cache in front of the map that remembers the most recently used endpoint is read and written unsynchronised on the hot path. The cache is the subtle half. It is also why Reset leaves a stale bucket behind, which is a visible bug rather than a detector warning. The tempting fix is one exclusive mutex around every method, which is race free, correct, and puts every exporter read in a queue behind the others: the hidden suite measures concurrent reads of a deliberately slow read path and rejects it.