For healthy == IGNORE, we return before the call to shard_next(), so
at this point we do want to consider health state - and it makes
absolutely no sense to give up a healthy backend in favor of an un-
healthy one anyway.

Fixes #2823