Add troubleshooting guide for timeseries with NaN

[sed-i]

Via @jameinel via matrix.

When timeseries have NaNs,

avg_over_time(
  juju_apiserver_request_duration_seconds{
    juju_controller="$controller",
    juju_unit=~"$controller_host",
    method="FullStatus",
    quantile="0.99",error_code!~"not found|unauthorized access"
  }[5m]
)

then aggregation operations produce unexpected results:

avg without (version, error_code) 
(
  juju_apiserver_request_duration_seconds{
    juju_controller="$controller",
    juju_unit=~"$controller_host",
    method="FullStatus",
    quantile="0.99",error_code!~"not found|unauthorized access"
  }[5m]
)

and only when we filter out the NaNs with e.g.

avg(avg_over_time((juju_apiserver_request_duration_seconds > 0.0001)[1h:]))

then we get the expected results.

We should document:

• How to avoid NaNs at instrumentation time.
• Treating potential NaN in dashboard/alert exprs.