Distributed deadlock detection improvement - Cleanup of backend pid association

As we've discussed in #1529 on [this](https://github.com/citusdata/citus/pull/1529#discussion_r130727361) comment, we should revisit the logic to associate backends after the deadlock is found.

We currently iterate on all backends to find the `pid`s of the backends that involve in the deadlock. We shouldn't need that, instead, we could store the `pid`s of initiating processes at the start. (Or something like that)