[casibbald]

Adds an optional op-timing feature. With it on, the server records its phases in op_duration_seconds{component="may_minihttp.server", op, outcome} (microscaler-timing). A host that renders that family shows them with no wiring.

op	from	to
accept_to_first_byte	connection accepted	its first request byte read (once per connection)
read_parse	a request's first byte	request fully parsed
service	handed to the service	service returned (err on failure)
response_write	service returned	response fully written to the socket
request_total	first byte	response fully written

Gauges, in op_gauge: open_connections, plus accept_backlog and accept_backlog_limit on Linux. The last two are read at scrape time from TCP_INFO on the listening socket; for a listener, tcpi_unacked and tcpi_sacked are the accept queue's length and limit.

Together these show where time goes that a client sees before the application's own timers start: accept queueing, a slow request read, or the response write.

The feature is off by default, and without it every hook is an empty inline function. With it, each request costs a few Instant::now() calls and atomic adds, and each connection allocates one small Vec.

Test: tests/server_timing.rs (--features op-timing) covers every op, a failed service call, the gauges, and open_connections returning to 0 after close. The default-feature server tests and the client lib tests also pass.