Make sure to run accept_enable when handling accept_backoff_ev

[agusdallalba]

This fixes issue #99.

There's a situation where the connection pressure is solved but redsocks doesn't resume listening to new connections.

1. redsocks_conn_max is hit, conn_pressure runs.
2. There are no long-idle connections, so conn_pressure adds the accept_backoff_ev event and disables all listeners.
3. Some time passes.
4. accept_backoff_ev gets triggered and redsocks_accept_backoff gets called.
5. conn_pressure_ongoing() == true, so conn_pressure runs.
6. This time there's a connection that's more than 7440s old. conn_pressure drops it and returns.
7. Now conn_pressure_ongoing() == false, but accept_enable isn't called because the conditional was evaluated before running conn_pressure.
8. redsocks_accept_backoff returns, so accept_backoff_ev is not pending anymore and accept_enable will never be called.

[darkk]

Thank you for the test-case!
The fix seems a bit questionable to me - enabling listeners on every exponential backoff iteration also looks like invariant violation, but the bug is there for sure.

[agusdallalba]

I could have added an accept_enable() in redsocks.c:1032, right before the return, but my reasoning was that the accept_backoff_ev event being pending "means" that redsocks isn't listening at the moment—and the rest of the code treats it that way. For example conn_pressure_lowered calls accept_enable() only if this event is pending.

So, in this way, event_add(&accept_backoff_ev, &tvdelay) semantically is 'stop listening for tvdelay'. If the event handler can return without re-enabling the listeners this meaning is lost. I chose to call accept_enable() as soon as possible within the event handler to prevent another bug like this one from ever happening.