If resume continued to run through a disconnect/reconnect
cycle then a race would be caused by c.stop = make(chan struct{})
Likely to be a rare issue in production (a lot of messages persisted,
slow store, and/or fast reconnect).
Also to enable tests to be run with the race detector the
Test_ConnectRetry has been updated to avoid a race.

Signed-off-by: Matt Brittan <matt@brittan.nz>

The initil implementation of this made client.go more
confusing than was required (as the change was just
to support testing). Added a mutex to support this as
that approach might be useful with other tests too.

Signed-off-by: Matt Brittan <matt@brittan.nz>