repeatedly releasing and leasing from the same thread results in starvation

[stchang]

Repeatedly releasing and leasing (from the same thread) results in starvation due to hitting the max number of threads because the old conns never finish getting returned (this is assuming #6 is fixed).

For example, the following test fails:

(check-not-exn
  (λ ()
    (let loop ([n 100])
      (unless (zero? n)
        (connection-pool-return p (connection-pool-lease p))
        (loop (sub1 n))))))

I'm guessing it's due to the slowness of unsubscribe because it works if I remove that cmd in release-conn. Maybe we need to treat publish/subscribe cmd as a special case.