As initially described in #730, there's a general need to understand the proxy's latency overhead. One could go pretty deep trying to answer this question, but I think we can start with a reasonably helpful indicator: the amount of time a request spends in the proxy.

I propose adding a new request_handle_µs histogram to the proxy with only a direction label. This histogram should store the elapsed time (in microseconds) from the moment a request reaches the source stack until the request is dropped.