Allow finer grained caching of metrics in metric_frequency

[AntPAllen]

At the moment the metric_frequency processor caches metrics just based on the metric name. This means that metrics that come from for example kube-state-metrics aren't accurately reported due to the metrics being differentiated on attributes/labels. A suggestion would be to be able to cache the metrics where the key is a struct that contains both the metric name and its attributes.

[JonnySparrenhok]

It would also be great to have the ability to exclude certain services from being dropped by metric_frequency. Some services may have metrics that are very important to scrape as often as possible while others are fine to drop by metric_frequency. It would be a very appreciated feature if you add the ability to only apply metric_frequency to defined jobs/namespaces/metrics.