Stop holding the write lock across proxy shutdown drain

[damilolaedwards]

Summary

Stop held the server's write lock for its entire duration, including the
call to httpSrv.Shutdown, which blocks until in-flight requests finish. The
/ready handler took the same lock to read the readiness flag, so a /ready
request landing while Stop held the lock would park on it, keeping its
connection open. Shutdown then waited on that same connection to go idle,
which could never happen while the handler waited on a lock only Stop's own
completion would release. A rolling restart with active readiness probing
could deadlock for the entire shutdown budget.

Makes the readiness flag an atomic so /ready and Ready() never contend for
the lock at all, and restructures Stop to hold the lock only for the
bounded teardown steps, releasing it before the blocking shutdown and drain
wait. Stop is also now idempotent via an atomic compare-and-swap instead of
a started check under the lock.

Test plan

• go test -race ./pkg/proxy/...
• New tests drive a real running server under concurrent /ready traffic
  and confirm Stop returns promptly, and confirm Stop is safe under
  concurrent callers; confirmed the traffic test fails (full timeout,
  deadline exceeded) on the old code
• go build ./..., go vet ./...