net/http/internal/http2: let the server's serve goroutine exit when idle

Implement the TODO from 2014 at the top of server.go: a server
connection with nothing to do used to pin a goroutine blocked in the
serve loop (reached through net/http's conn.serve and ServeConn) in
addition to the readFrames goroutine; now it pins only readFrames.

When the serve loop has nothing runnable (nothing being written,
unflushed, or poppable on the write side, and no shutdown in
progress), the serve goroutine parks: it simply exits. Every sender
into the serve loop's channels first calls beginServeSend, which
counts in-flight sends and revives a parked loop on a new goroutine,
so a frame from the peer, a handler write or body read, a timer
firing, or a graceful shutdown request wakes the connection back up.

Open streams and running handlers don't prevent parking, since the
handler-side paths into the serve loop (response writes, body reads,
stream deadline timers, pushes) all go through beginServeSend too.
This helps SSE/long-poll-style servers: a connection whose handler is
parked mid-long-poll for hours pins only the readFrames and handler
goroutines, and the handler's next Write or Flush revives the loop.
With open streams, the write scheduler can still hold DATA frames
blocked on flow control; that's fine to park on, because only an
incoming WINDOW_UPDATE or SETTINGS frame can unblock them, and
incoming frames revive the loop.

The serve loop's teardown became explicit rather than deferred, and
it takes over the cleanup that ServeConn's callers used to do after
it returned (closing the connection, unregistering it, and running
net/http's StateClosed hook), signaled through the new
ServeConnOpts.OnClose hook. Callers that don't set OnClose keep the
old blocking ServeConn semantics and never park.

Parking happens between the frames of an active request, so the
park/resume cycle runs on hot paths. Spawning the resume goroutine
with a pre-allocated method-value closure keeps it allocation-free:
BenchmarkClientServer/h2 reports the same 68 allocs/op as before,
with ns/op unchanged within noise.

As a temporary safety measure, the GODEBUG=http2serveparking=0
setting disables parking, restoring the old behavior of keeping a
goroutine parked for the lifetime of each connection. The setting is
undocumented (#-prefixed), so it is not listed in doc/godebug.md or
runtime/metrics.

This is the server-side counterpart of CL 810780, which let the
Transport's request-write goroutine exit early for the same reason:
for servers with many mostly-idle connections, a parked goroutine and
its stack per connection add up.

Updates #80735
Updates #81524

Change-Id: I4a9fd4e2ce7a7a4e0f84179e54a9cc97d1a62ad5
Reviewed-on: https://go-review.googlesource.com/c/go/+/831084
Reviewed-by: Nicholas Husin <husin@google.com>
Reviewed-by: Nicholas Husin <nsh@golang.org>
LUCI-TryBot-Result: golang-scoped@luci-project-accounts.iam.gserviceaccount.com <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Damien Neil <dneil@google.com>