this reverts a change made for #374, which inadvertently disabled inactivity timeouts altogether. instead we always QueueHeartbeat when bConnectPending changes, so GetNextHeartbeat can calculate a timeout based on the new connection state. i confirmed this fixes the broken test on master, and the original "20s delay" bug (using the repro in #393). this should also address the timeout stacking issue reported in igrigorik/em-http-request#149, as well as the original report in igrigorik/em-http-request#222
…begin called on the connection when epoll is enabled. Fix this by only setting EPOLLIN in the new ConnectionDescriptor if SelectForRead is true. If pause is called in post_init, SelectForRead will have become false during EventCallback.