Take only overload into account which is within the past 18h

We realized that overload monitoring taking in particular the 72h timeframe for overload-general into account is leading potentially to a lot of noise (see: analysis#27 (closed) for details). We tested with an 18h timeframe instead under the assumption that a) a new descriptor gets published at least every 18h and b) a sustained overload would give an updated timestamp in each of those newly uploaded descriptors. The resulting data looks much better for overload monitoring. So, let's in general just count overload that happened in the last 18h as overload relevant for network-health monitoring purposes.