Not instantiating an exporter and logging instead of failing to start
when we can't connect to one of the configured clusters is not the
behaviour we want; a user may think that ceph_exporter is monitoring all
of the requested clusters (since it's up and responds to metrics
requests) when in fact data is missing.
- `ceph.Conn` interface no longer depends on go-ceph/rados,
now defines its own `PoolStat` structure for our use.
- New separate `rados` package that implements the interface
- Merged `mocks` package into `ceph` to avoid circular import