Right now, we spawn potentially several connections without necessarily
needing to. For example, if we're transferring two objects, we can
practically use at most three connections: one for the batch and one for
each object.
To make this more efficient and avoid needless overhead, let's not
actually create the connection until we attempt to acquire it. At that
point, if it doesn't exist, we'll spawn a new one (using the control
path socket if possible) and then start sending data. Otherwise, unless
it's the initial connection, let's just stub it out until we actually
need it.
Note that we still create the connection before it's needed because we
create all workers up front, but that will change in a future commit now
that we have this change in place.
Co-authored-by: Chris Darroch <chris8088@github.com>
44b8801cbb