Better Async Processing

Currently every connection uses it's own thread from a thread pool.
This is fine, but not optimal and could be enhanced to run multiple
connections per thread.

This is already 50% implemented with the new epoll() functionality, I think.

Start by reviewing the current per-connection thread-pool handling and the existing epoll() functionality mentioned in the issue. Determine how connections are assigned to threads and what remains incomplete. Done means multiple connections can share a thread without breaking the transport layer behavior.

Written by the indexing model from the issue text.