Fix for HTTP2 request hanging bug.
If, when a socket request completed asynchronously, the next
socket request in the queue failed synchronously,
ClientSocketPoolBase wouldn't continue to try to service other
socket requests in the queue.
As a result, the socket pools could end up with pending connect
requests and free socket slots, but no ConnectJobs would be made
to service them, if this happened 6 times in a row to a socket
This was a relatively obscure issue, until H2 started depending
on this path for sharing sessions when different domains map to
the same IP.