open connections in a way resilient to temporary failures

open connections in a way resilient to temporary failures

In case of connection failure (e.g. a backend getting terminated by
OOM killer and taking down the whole server), the workers failed
with an exception without putting a result into the queue, leading
to hangs.

This fixes that by making the workers resilient to connection issues
by catching the exceptions and terminating nicely. Also, we retry
connection several times, to continue with benchmarking even after
OOM hits us unexpectedly, etc.