Previously, both ends were set to non-blocking I/O.  Now, only the
reader is set non-blocking.  This means the writer will block for a while
if the pipe fills, but since the watcher is in a different thread from
the reader, the pipe will eventually drain, and the writer can continue.

This is really nothing more than a lock, just not a mutex-type lock.

EHOSTUNREACH

The manager will no longer try to dispatch internal events for dead sockets,
preventing a situation where the ref count could go from 0 back to 1!

retrying on a socket is pointless, as it will immediately fail.

would be so nice if people would define 'msg_control' in a msghdr to be
consistant with other pointers...

statements.

flag for us.  To see if a UDP packet was truncated, look at
done_event->attributes and see if ISC_SOCKEVENTATTR_TRUNC is set.

requiring an isc_region_t to fill in.  Use these macros in a few places.

routines (internal and external)