After this change, the library got issues with stopping cluster properly: https://github.com/couchbaselabs/couchbase-cxx-client/pull/359/files#diff-d57958cc60b900a216af3a607194ccac57e75457e69eddfa953cca5b33f2cc78R59.

So I dig it a bit and found several issues:

• meter uses raw pointer to capture itself in the periodic timer (which causes leak and potential use-after-free)
• we should really try to signal timers in tracer and meter before relying on destructors, because some references might be still hold by IO threads
• when error occurs during core::cluster::open(), we should ensure that the cluster in closed state, in particular meter/tracer should be also stopped.