in vcl_backend_response {} and vcl_backend_error {} we test if
`bereq.backend == vk8s_cluster.backend(resolve=LAZY)`. With the
current code, this condition will never evaluate true, because
`resolve=NOW` will make the shard director return the "real" (VBE)
backend.

With this patch, the net effect, the chosen cluster varnish backend,
should still be the same, except for the rare race event that the
backend becomes unhealthy between the time the vcl code executes and
when the backend connection is made. We already account for this race
by setting `req.hash_ignore_busy` in vcl_revc {}.